# PostTrainLLM

> A Mac-local LLM specialist factory: post-training, evidence-gated packaging,
> MLX runtime work, and a WebGPU playground.

## When to use this

Reach for PostTrainLLM when you need to train, fine-tune, evaluate, or package a specialist LLM on a single Apple Silicon Mac — without cloud compute. Best-fit jobs:

- Fine-tuning an open model with LoRA/QLoRA on a Mac (MLX)
- Evaluating a local model against frozen benchmarks (BFCL, tool-calling, perplexity)
- Packaging a trained specialist for MLX or on-device inference
- Comparing Mac-local training approaches (distillation, GRPO, SFT)
- Training a tiny model in-browser and inspecting its loss and samples
- Running client-side inference and validating WebGPU kernels

Do not use PostTrainLLM for: large-scale distributed training, frontier-scale pretraining, or anything that requires a GPU cluster — it is designed for one Mac.

## Public surface

- [Home](https://posttrainllm.com/): Product and research-lab overview
- [Native factory quickstart](https://posttrainllm.com/docs/quickstart): Target → data → post-train → eval → package → report
- [Browser training lab](https://posttrainllm.com/playground): TinyGPT training with visible loss and samples
- [Browser inference lab](https://posttrainllm.com/inference): Client-side model inference
- [WebGPU kernel tests](https://posttrainllm.com/webgpu-test): Browser kernel correctness and capability probes
- [Training dashboard](https://posttrainllm.com/training-dashboard): Inspect recorded training state
- [Experiments](https://posttrainllm.com/experiments): 127 resolved attempts — 5 worked, 55 worked with caveat, 67 non-positive or mixed
- [Recipes](https://posttrainllm.com/recipes): 18 reproducible or explicitly bounded recipe contracts
- [Learning paths](https://posttrainllm.com/learn): 9 paths and 13 buildable artifacts
- [Studied projects](https://posttrainllm.com/studies): 20 source-linked evidence dossiers
- [Inspiration](https://posttrainllm.com/inspiration): credited tools and projects, local lessons, and source trails
- [Documentation](https://posttrainllm.com/docs/): 324 source documents
- [Artifacts](https://posttrainllm.com/artifacts): Public packages, evidence, and blockers
- [Benchmarks](https://posttrainllm.com/benchmarks): Measured performance evidence
- [Evaluation leaderboard](https://posttrainllm.com/leaderboard): Model-quality comparisons and gates
- [Mac app](https://posttrainllm.com/download): Verified release 0.2.0 (2)
- [Devlog](https://posttrainllm.com/devlog): Build history
- [Agent catalog](https://posttrainllm.com/api/ai): Complete page-to-Markdown inventory
- [OpenAPI spec](https://posttrainllm.com/openapi.json): Machine-readable API description
- [Sitemap](https://posttrainllm.com/sitemap.xml): Canonical public HTML routes
- [Full agent index](https://posttrainllm.com/llms-full.txt): Every public page grouped by kind

## Boundaries

Public machine resources are catalogued separately from HTML pages. Local
factory runs, models, private artifacts, and unpublished evidence are excluded.
`decision.json` remains the terminal quality and product authority.

## CLI

PostTrainLLM ships a native Mac CLI for the full factory loop:

```bash
# build the native factory CLI
git clone https://github.com/PostTrainLLM/posttrainllm && cd native-mac
swift build --product posttrainllm

# distill a specialist, gate it, serve it
posttrainllm distill --teacher qwen3 --student …
posttrainllm eval-gate --spec sql.json --candidate …
posttrainllm serve --port 8080  # OpenAI-compatible
```

The CLI exposes 100+ subcommands covering train, eval, package, serve, and
inspect. See [the generated CLI reference](https://posttrainllm.com/docs/cli-reference) or
`posttrainllm --help` for the full command surface.

## Application and research pages

- [PostTrainLLM — Build and measure specialist LLMs on a Mac](https://posttrainllm.com/) — [Markdown](https://posttrainllm.com/index.md) — Build bounded specialist language models sized to one Apple Silicon machine, then distill, fine-tune, evaluate, package, and inspect them with reproducible…
- [How to build an agent trajectory dataset | PTLLM](https://posttrainllm.com/agent-trajectory-dataset) — [Markdown](https://posttrainllm.com/agent-trajectory-dataset.md) — Build an agent trajectory dataset with typed events, provenance, privacy filters, leakage-safe splits, and outcome-based evaluation.
- [Apple Foundation Models on-device: measured capabilities and limits | PTLLM](https://posttrainllm.com/apple-foundation-models) — [Markdown](https://posttrainllm.com/apple-foundation-models.md) — We scored Apple's on-device Foundation Models framework on real agentic gates. Measured verdict: a free, private floor for lightweight turns — it cannot…
- [posttrainllm — articles](https://posttrainllm.com/articles) — [Markdown](https://posttrainllm.com/articles.md) — Interactive field guides and long-form articles on model development, Mac-local post-training, evaluation, and hardware limits.
- [Measuring Model Capability Per Gigabyte of RAM: Beyond Parameter Chasing | PTLLM](https://posttrainllm.com/articles/capability-per-gigabyte-ram-small-llms) — [Markdown](https://posttrainllm.com/articles/capability-per-gigabyte-ram-small-llms.md) — Stop chasing model size. Learn how to evaluate local LLMs using capability per GB of RAM, active parameter efficiency, and hardware-aware density metrics.
- [How to compare two fine-tuning runs without moving the evaluation target | PTLLM](https://posttrainllm.com/articles/how-to-compare-two-fine-tuning-runs-without-moving-the-evaluation-target) — [Markdown](https://posttrainllm.com/articles/how-to-compare-two-fine-tuning-runs-without-moving-the-evaluation-target.md) — Learn how to compare LLM fine-tuning runs effectively by freezing baselines, using frontier-calibrated gates, and measuring both primary and regression scores.
- [How to cook an LLM | posttrainllm](https://posttrainllm.com/articles/how-to-cook-an-llm) — [Markdown](https://posttrainllm.com/articles/how-to-cook-an-llm.md) — A hands-on journey from garden to table: gather data, inspect tokens, clean a batch, explore pre-training, then refine and serve. Kitchen and model-building…
- [A Practical LoRA Evaluation Workflow for Mac-Local Models | PTLLM](https://posttrainllm.com/articles/lora-evaluation-workflow-mac-local-models) — [Markdown](https://posttrainllm.com/articles/lora-evaluation-workflow-mac-local-models.md) — Learn how to evaluate Mac-local LoRA adapters using frozen baselines, frontier calibration, AST matching, regression gates, and automated report cards.
- [How to Post-Train a Small Language Model on Apple Silicon | PTLLM](https://posttrainllm.com/articles/post-train-small-language-model-apple-silicon) — [Markdown](https://posttrainllm.com/articles/post-train-small-language-model-apple-silicon.md) — Learn how to post-train 0.6B–4B language models locally on Apple Silicon using LoRA, MPS, and MLX. Includes real run data, factory loops, and hardware…
- [posttrainllm — public artifacts](https://posttrainllm.com/artifacts) — [Markdown](https://posttrainllm.com/artifacts.md) — Public posttrainllm artifacts: measured model packages, routed specialists, benchmark reports, blockers, and release status.
- [Artifact: ANE M8 Core ML Chain | PTLLM](https://posttrainllm.com/artifacts/ane-m8-coreml-chain) — [Markdown](https://posttrainllm.com/artifacts/ane-m8-coreml-chain.md) — A layer-chunked Core ML chain ran a Qwen3 28-block path on the Apple Neural Engine at about 17 tok/s.
- [Artifact: Browser WebGPU Training Speedup | PTLLM](https://posttrainllm.com/artifacts/browser-webgpu-speedup) — [Markdown](https://posttrainllm.com/artifacts/browser-webgpu-speedup.md) — A frozen ABBA run on an Apple M5 Pro measured a 10.67x Large-preset WebGPU speedup over WASM with 4.72% paired final-loss drift and zero runtime errors.
- [Artifact: Factory Run Schema v1 | PTLLM](https://posttrainllm.com/artifacts/factory-run-schema-v1) — [Markdown](https://posttrainllm.com/artifacts/factory-run-schema-v1.md) — The canonical target, data, post-training, evaluation, package, and report shape for PostTrainLLM runs.
- [Artifact: 4-bit Browser Gallery Models | PTLLM](https://posttrainllm.com/artifacts/gallery-int4-browser-models) — [Markdown](https://posttrainllm.com/artifacts/gallery-int4-browser-models.md) — The browser gallery ships fp16 and int4 variants so model downloads are smaller and cold-start is cheaper.
- [Artifact: Hugging Face Specialist Model Archive v1 | PTLLM](https://posttrainllm.com/artifacts/hf-specialist-model-archive-v1) — [Markdown](https://posttrainllm.com/artifacts/hf-specialist-model-archive-v1.md) — The local specialist model cache was promoted to Hugging Face or deleted when safely re-downloadable from upstream repos.
- [Artifact: Huge Preset Decode Throughput | PTLLM](https://posttrainllm.com/artifacts/huge-decode-throughput) — [Markdown](https://posttrainllm.com/artifacts/huge-decode-throughput.md) — The native Mac runtime reached high local decode throughput on the Huge preset, showing the serving path is viable for local eval loops.
- [Artifact: Memory64 Browser Behemoth Allocation | PTLLM](https://posttrainllm.com/artifacts/memory64-browser-behemoth) — [Markdown](https://posttrainllm.com/artifacts/memory64-browser-behemoth.md) — A WebAssembly Memory64 build lifted the browser model allocation ceiling past the old 4GB tab limit.
- [Artifact: Needle 2: The 45M Capacity Boundary | PTLLM](https://posttrainllm.com/artifacts/needle2-tool-selection) — [Markdown](https://posttrainllm.com/artifacts/needle2-tool-selection.md) — Catalog restriction and a four-arm successor factorial both failed, establishing that this 45M model class—not one missing recipe tweak—is the current…
- [Artifact: OffHours Context-Interference Benchmark | PTLLM](https://posttrainllm.com/artifacts/offhours-context-interference) — [Markdown](https://posttrainllm.com/artifacts/offhours-context-interference.md) — A paired office-work benchmark that separates unresolved personal context from interruption and raw context volume. Devin showed no mental-toll penalty, but…
- [Artifact: Pace Intent Router v8 | PTLLM](https://posttrainllm.com/artifacts/pace-intent-router-v8) — [Markdown](https://posttrainllm.com/artifacts/pace-intent-router-v8.md) — A 49.5M intent classifier that was extremely fast and strong on its source-matched synthetic holdout, then failed the first leakage-checked sealed…
- [Artifact: Parakeet WGSL Browser ASR | PTLLM](https://posttrainllm.com/artifacts/parakeet-wgsl-browser-asr) — [Markdown](https://posttrainllm.com/artifacts/parakeet-wgsl-browser-asr.md) — Parakeet v3 scored 0/82 word errors and ran 3.51x faster than native WhisperKit on the same eight LibriSpeech clips, while missing a deliberately aggressive…
- [Artifact: Qwen3-0.6B Routed SQL Specialist | PTLLM](https://posttrainllm.com/artifacts/qwen06-sql-routed-v1) — [Markdown](https://posttrainllm.com/artifacts/qwen06-sql-routed-v1.md) — A two-adapter routed SQL artifact: public schema-only SQL routes to a b-mc2 adapter; local SQLite execution routes to a synthetic execution adapter.
- [Artifact: Qwen3-4B File-Ops Distilled | PTLLM](https://posttrainllm.com/artifacts/qwen3-4b-file-ops-distilled) — [Markdown](https://posttrainllm.com/artifacts/qwen3-4b-file-ops-distilled.md) — A routed 4B file-operation specialist distilled from frontier/gold trajectories, with the breadth regression disclosed in the package.
- [Artifact: Qwen3-4B Multibackend Distilled | PTLLM](https://posttrainllm.com/artifacts/qwen3-4b-multibackend-distilled) — [Markdown](https://posttrainllm.com/artifacts/qwen3-4b-multibackend-distilled.md) — A public failed attempt that retained 100% file-ops depth but drove the recorded breadth score down to 31%.
- [Artifact: Qwen3-4B ReST Fused | PTLLM](https://posttrainllm.com/artifacts/qwen3-4b-rest-fused) — [Markdown](https://posttrainllm.com/artifacts/qwen3-4b-rest-fused.md) — Fresh requalification confirms a perfect, safer, faster file-ops specialist while rejecting it as the general successor because breadth regressed.
- [Artifact: VibeThinker-3B Agentic Distilled | PTLLM](https://posttrainllm.com/artifacts/vibethinker-3b-agentic-distilled) — [Markdown](https://posttrainllm.com/artifacts/vibethinker-3b-agentic-distilled.md) — A preserved agentic distillation checkpoint whose public weights outlived its evaluation evidence; no current win can be claimed.
- [Artifact: VibeThinker-3B MLX Conversion | PTLLM](https://posttrainllm.com/artifacts/vibethinker-3b-mlx) — [Markdown](https://posttrainllm.com/artifacts/vibethinker-3b-mlx.md) — A public Apple-Silicon-friendly conversion of WeiboAI/VibeThinker-3B, preserved as a runtime artifact rather than claimed as a PostTrainLLM-trained model.
- [Reproducible LLM benchmarks — posttrainllm](https://posttrainllm.com/benchmarks) — [Markdown](https://posttrainllm.com/benchmarks.md) — Browse reproducible language-model benchmarks with frozen protocols, prerecorded outputs, local replay commands, and specialist comparisons.
- [Cross-game model arena — posttrainllm](https://posttrainllm.com/benchmarks/arena) — [Markdown](https://posttrainllm.com/benchmarks/arena.md) — Inspect provisional Arena Elo for head-to-head games and paired score evidence for single-player games without conflating the two.
- [Character Chess LLM benchmark — posttrainllm](https://posttrainllm.com/benchmarks/chess) — [Markdown](https://posttrainllm.com/benchmarks/chess.md) — Watch frontier, 9B, and 4B language models solve identical chess tactics and replay their complete paired games move by move.
- [Character 2048 LLM benchmark — posttrainllm](https://posttrainllm.com/benchmarks/game-2048) — [Markdown](https://posttrainllm.com/benchmarks/game-2048.md) — Inspect the retained Character 2048 benchmark failure, replay every local model move, and see why the expected intelligence advantage did not reproduce.
- [How to build a small language model specialist | PTLLM](https://posttrainllm.com/build-small-language-model-specialist) — [Markdown](https://posttrainllm.com/build-small-language-model-specialist.md) — Build a routed small language model specialist with a frozen baseline, narrow training objective, regression gates, packaging, and an honest decision.
- [posttrainllm — changelog](https://posttrainllm.com/changelog) — [Markdown](https://posttrainllm.com/changelog.md) — Verified product improvements shipped by the posttrainllm Mac-local specialist factory.
- [posttrainllm — devlog](https://posttrainllm.com/devlog) — [Markdown](https://posttrainllm.com/devlog.md) — Notes from building posttrainllm — kernel measurements, honest negative results, decisions made while AI-pairing.
- [PostTrainLLM for Mac — release status](https://posttrainllm.com/download) — [Markdown](https://posttrainllm.com/download.md) — Verified release status for the native PostTrainLLM app on Apple Silicon. The signed, notarized, stapled DMG is published with its SHA-256.
- [posttrainllm — eval leaderboard](https://posttrainllm.com/eval-leaderboard) — [Markdown](https://posttrainllm.com/eval-leaderboard.md) — Drop E0 eval JSONL files to compare models, checkpoints, and tasks in the browser.
- [How to evaluate a local LLM | PTLLM](https://posttrainllm.com/evaluate-local-llm) — [Markdown](https://posttrainllm.com/evaluate-local-llm.md) — Evaluate a local LLM with paired target and regression checks. Walk through a real report card and validate its evidence format offline without loading a model.
- [Every experiment — posttrainllm](https://posttrainllm.com/experiments) — [Markdown](https://posttrainllm.com/experiments.md) — The complete PostTrainLLM experiment archive: every worked, failed, regressed, inconclusive, superseded, and rejected attempt with evidence and a reusable…
- [How to train and fine-tune your own LLM on a Mac | PTLLM](https://posttrainllm.com/fine-tune-llm-on-mac) — [Markdown](https://posttrainllm.com/fine-tune-llm-on-mac.md) — Practical guide to train your own specialist LLM on Apple Silicon: define a target, LoRA/QLoRA post-train, evaluate against a frozen baseline, and package…
- [How to train your own LLM (without a GPU cluster) | PTLLM](https://posttrainllm.com/how-to-train-your-own-llm) — [Markdown](https://posttrainllm.com/how-to-train-your-own-llm.md) — What “train your own LLM” usually means in 2026: post-train a base model with LoRA/QLoRA, evaluate honestly, and ship a specialist. Includes Mac-local paths…
- [posttrainllm — in-browser inference](https://posttrainllm.com/inference) — [Markdown](https://posttrainllm.com/inference.md) — Run real LLMs fully client-side in your browser via WebGPU (transformers.js). Pull any ONNX model from Hugging Face by id, or pick a curated one — no…
- [Inspiration: the projects behind PostTrainLLM](https://posttrainllm.com/inspiration) — [Markdown](https://posttrainllm.com/inspiration.md) — A public thank-you to the tools and projects that shaped PostTrainLLM, with source links, local lessons, evidence, and clear boundaries.
- [posttrainllm — leaderboard](https://posttrainllm.com/leaderboard) — [Markdown](https://posttrainllm.com/leaderboard.md) — Browser-trained, browser-runnable LLMs ranked on small benchmarks. Pareto views by params and by training compute.
- [Mac-local LLM specialist factory | PostTrainLLM](https://posttrainllm.com/learn) — [Markdown](https://posttrainllm.com/learn.md) — Follow the source-linked loop from target and data through post-training, evaluation, packaging, and an honest ship, retry, or reject report.
- [Learn language models from first principles | PostTrainLLM](https://posttrainllm.com/learn/session) — [Markdown](https://posttrainllm.com/learn/session.md) — Learn parameters, gradients, tensors, transformers, post-training, and evaluation through public lessons, browser-local exercises, and delayed recall.
- [MLX LoRA fine-tuning on Apple Silicon | PTLLM](https://posttrainllm.com/mlx-lora-fine-tuning) — [Markdown](https://posttrainllm.com/mlx-lora-fine-tuning.md) — Learn the MLX LoRA and QLoRA workflow on Apple Silicon, from JSONL data and bounded training to evaluation, fusion, and evidence.
- [posttrainllm — train a real transformer in your browser tab](https://posttrainllm.com/playground) — [Markdown](https://posttrainllm.com/playground.md) — Train a real GPT-2 from scratch in your browser tab with hand-written, parity-tested WASM and WebGPU backends. A paired Apple M5 Pro receipt verifies 10.67x…
- [PostTrainLLM vs Axolotl for LLM fine-tuning | PTLLM](https://posttrainllm.com/posttrainllm-vs-axolotl) — [Markdown](https://posttrainllm.com/posttrainllm-vs-axolotl.md) — Compare PostTrainLLM's Apple Silicon specialist-model workflow with Axolotl's configuration-driven LLM fine-tuning framework.
- [PostTrainLLM vs Unsloth for local LLM fine-tuning | PTLLM](https://posttrainllm.com/posttrainllm-vs-unsloth) — [Markdown](https://posttrainllm.com/posttrainllm-vs-unsloth.md) — Compare PostTrainLLM's Apple Silicon specialist-model factory with Unsloth's optimized LLM training and inference framework.
- [posttrainllm — recipes](https://posttrainllm.com/recipes) — [Markdown](https://posttrainllm.com/recipes.md) — The recipe log: named post-training and packaging recipes for posttrainllm. Each names a failure mode, its data, an eval gate, and a stop rule — not just…
- [posttrainllm — the performance journey](https://posttrainllm.com/roadmap) — [Markdown](https://posttrainllm.com/roadmap.md) — The performance journey of posttrainllm — every speed lever, what's shipped, what's blocked, and why.
- [posttrainllm — SAE timeline](https://posttrainllm.com/sae-timeline) — [Markdown](https://posttrainllm.com/sae-timeline.md) — Drop a posttrainllm SAE timeline JSONL to see reconstruction error and sparse-feature activity over checkpoints.
- [posttrainllm — verified WebGPU speed on one Mac](https://posttrainllm.com/speedup) — [Markdown](https://posttrainllm.com/speedup.md) — A paired Apple M5 Pro WebGPU-versus-WASM receipt verifies 10.67x on the Large preset; the older multi-preset curve remains historical.
- [Studied projects and research | PostTrainLLM](https://posttrainllm.com/studies) — [Markdown](https://posttrainllm.com/studies.md) — Canonical, source-linked evidence dossiers for every external product, model, platform, technical writeup, and industry case study retained by PostTrainLLM.
- [posttrainllm — training run dashboard](https://posttrainllm.com/training-dashboard) — [Markdown](https://posttrainllm.com/training-dashboard.md) — Drop a .jsonl training log to see loss, LR, val, step-rate, and spike markers — overlay multiple runs to compare.
- [posttrainllm — WebGPU kernel tests](https://posttrainllm.com/webgpu-test) — [Markdown](https://posttrainllm.com/webgpu-test.md) — Mac-local LLM specialist factory — post-training and runtime that fits on one Mac, plus WebGPU playground.

## Studied products and research

- [Apple on-device Foundation Models | PostTrainLLM](https://posttrainllm.com/studies/apple-on-device-foundation-models) — [Markdown](https://posttrainllm.com/studies/apple-on-device-foundation-models.md) — Apple's system model is a free and private routing floor, but its context and action-grounding limits make it unsuitable as the core capability dependency…
- [Baseten post-training positioning | PostTrainLLM](https://posttrainllm.com/studies/baseten-post-training) — [Markdown](https://posttrainllm.com/studies/baseten-post-training.md) — Post-training is a system of custom data, reward shaping, performance work, infrastructure, evaluation, and packaging rather than a generic fine-tuning…
- [Bonsai 2 27B | PostTrainLLM](https://posttrainllm.com/studies/bonsai-2-27b) — [Markdown](https://posttrainllm.com/studies/bonsai-2-27b.md) — Ternary representation, packing overhead, activation transforms, kernel support, artifact size, runtime memory, and capability retention can rank a model…
- [Castform RL fine-tune platform | PostTrainLLM](https://posttrainllm.com/studies/castform-rl-finetune) — [Markdown](https://posttrainllm.com/studies/castform-rl-finetune.md) — Composite rewards, trace-driven data loops, reasoning-depth classification, and explicit environment contracts make reinforcement-style fine-tuning more…
- [Cline and agent context hierarchy research | PTLLM](https://posttrainllm.com/studies/cline-context-hierarchy) — [Markdown](https://posttrainllm.com/studies/cline-context-hierarchy.md) — Structured-output enforcement and an explicit context hierarchy can reduce ambiguous agent actions and make tool boundaries easier to inspect. The research…
- [Defog SQLCoder and Arctic Text2SQL class models | PTLLM](https://posttrainllm.com/studies/defog-sqlcoder-arctic-text2sql) — [Markdown](https://posttrainllm.com/studies/defog-sqlcoder-arctic-text2sql.md) — Strong SQL systems report execution accuracy on substantial SQL benchmarks rather than relying on exact string match alone. External category scores remain…
- [Gigatoken | PostTrainLLM](https://posttrainllm.com/studies/gigatoken) — [Markdown](https://posttrainllm.com/studies/gigatoken.md) — SIMD and cache-heavy bulk BPE can make first-pass tokenization much faster on large corpora, including Qwen tokenizers on Apple Silicon. No current…
- [Halo / White Circle | PostTrainLLM](https://posttrainllm.com/studies/halo) — [Markdown](https://posttrainllm.com/studies/halo.md) — HF-native training and portable checkpoints, independent expert/context/tensor/data parallelism, faithful multi-turn trajectories, operation-specific model…
- [Hugging Face model hub and SQL specialists | PTLLM](https://posttrainllm.com/studies/hugging-face-sql-specialists) — [Markdown](https://posttrainllm.com/studies/hugging-face-sql-specialists.md) — Small public SQL models provide realistic baselines and expose how much benchmark, model size, and serving format affect apparently similar claims.…
- [Inside vLLM | PostTrainLLM](https://posttrainllm.com/studies/inside-vllm) — [Markdown](https://posttrainllm.com/studies/inside-vllm.md) — Admission control, scheduling, paged KV allocation, prefill, decode, continuous batching, caching, and distributed execution explain why a serving engine…
- [Needle 2 | PostTrainLLM](https://posttrainllm.com/studies/needle-2) — [Markdown](https://posttrainllm.com/studies/needle-2.md) — A compact call-only model combines top-five tool retrieval, constrained decoding, confidence escalation, bounded context, quantization, and a small packaged…
- [parakeet.wgsl | PostTrainLLM](https://posttrainllm.com/studies/parakeet-wgsl) — [Markdown](https://posttrainllm.com/studies/parakeet-wgsl.md) — Raw WebGPU plus SIMD-WASM can run Parakeet TDT browser inference with public kernels, a package format, cache behavior, a converter, and a benchmark…
- [QORL | PostTrainLLM](https://posttrainllm.com/studies/qorl) — [Markdown](https://posttrainllm.com/studies/qorl.md) — Parameter-aware query optimization needs equal search budgets, fresh held-out measurement, SQL-equivalence checks, and a native-planner baseline before an…
- [Savante and Aryabhata | PostTrainLLM](https://posttrainllm.com/studies/savante-aryabhata) — [Markdown](https://posttrainllm.com/studies/savante-aryabhata.md) — Curated domain data, filtered reasoning traces, supervised fine-tuning, model merging, and verifiable-reward reinforcement learning combine in a JEE…
- [Splash model-specific Mac inference | PostTrainLLM](https://posttrainllm.com/studies/splash-mac-inference) — [Markdown](https://posttrainllm.com/studies/splash-mac-inference.md) — Shape-specific Metal kernels, a model-specific speculative draft, packed weights, cache reuse, and a startup memory plan trade generic model support for…
- [TrainLoop AI | PostTrainLLM](https://posttrainllm.com/studies/trainloop-ai) — [Markdown](https://posttrainllm.com/studies/trainloop-ai.md) — Failed-attempt accounting, slice metrics, trace review, candidate selection, batch-first post-training, policy lag, and LoRA update geometry make…
- [TrainLoop LoRA geometry writeup | PostTrainLLM](https://posttrainllm.com/studies/trainloop-lora-geometry) — [Markdown](https://posttrainllm.com/studies/trainloop-lora-geometry.md) — Effective update geometry, layer placement, and controlled rank analysis can explain why two adapters with similar parameter counts behave differently.…
- [TrainLoop Mercor coding-agent post | PostTrainLLM](https://posttrainllm.com/studies/trainloop-mercor-coding-agent) — [Markdown](https://posttrainllm.com/studies/trainloop-mercor-coding-agent.md) — File-native and code-native harnesses may fit knowledge-work agents better than bespoke tool APIs because the environment already exposes durable state and…
- [TrainLoop NollaMD case study | PostTrainLLM](https://posttrainllm.com/studies/trainloop-nollamd) — [Markdown](https://posttrainllm.com/studies/trainloop-nollamd.md) — Candidate-selection framing can make a sparse task measurable when unconstrained generation is too difficult or too ambiguous to grade reliably. The framing…
- [TrainLoop OAPL writeup | PostTrainLLM](https://posttrainllm.com/studies/trainloop-oapl) — [Markdown](https://posttrainllm.com/studies/trainloop-oapl.md) — Batch rollouts, offline scoring, and compact updates can make a post-training loop easier to inspect, while policy lag can stabilize learning under some…

## Experiment records

- [A1 first tool-calling specialist harness | PTLLM](https://posttrainllm.com/experiments/a1-tool-caller-harness) — [Markdown](https://posttrainllm.com/experiments/a1-tool-caller-harness.md) — Reusable infrastructure can be superseded by stronger downstream evidence even when its first named model run never happened. Retained outcome: superseded.
- [Apple on-device Foundation Models action-grounding probe | PTLLM](https://posttrainllm.com/experiments/apple-foundation-models-action-grounding) — [Markdown](https://posttrainllm.com/experiments/apple-foundation-models-action-grounding.md) — Apple Foundation Models are useful as a free routing floor and intel source, not as the core capability dependency. Retained outcome: failed.
- [Autocorrect FLAN-T5-small ordinary-loss pilot (2026-07-25) | PTLLM](https://posttrainllm.com/experiments/autocorrect-ordinary-loss-pilot) — [Markdown](https://posttrainllm.com/experiments/autocorrect-ordinary-loss-pilot.md) — Minimum-edit repair is a narrow band between copying everything and rewriting everything Retained outcome: regressed.
- [Autocorrect FLAN-T5-small tiny-overfit gate (2026-07-25) | PTLLM](https://posttrainllm.com/experiments/autocorrect-tiny-overfit-gate) — [Markdown](https://posttrainllm.com/experiments/autocorrect-tiny-overfit-gate.md) — A memorization gate needs more than one distinct target Retained outcome: worked with caveat.
- [B1 SQL eval infrastructure | PostTrainLLM](https://posttrainllm.com/experiments/b1-sql-eval-infra) — [Markdown](https://posttrainllm.com/experiments/b1-sql-eval-infra.md) — Execution accuracy over SQLite is the right reusable gate; infra alone is not a specialist result. Retained outcome: worked with caveat.
- [Browser default corpus fix | PostTrainLLM](https://posttrainllm.com/experiments/browser-default-corpus-fix) — [Markdown](https://posttrainllm.com/experiments/browser-default-corpus-fix.md) — Default data is part of the product claim; a training demo's default corpus must produce non-memorized samples. Retained outcome: worked with caveat.
- [Browser generation streaming and KV cache | PTLLM](https://posttrainllm.com/experiments/browser-generation-streaming-kv-cache) — [Markdown](https://posttrainllm.com/experiments/browser-generation-streaming-kv-cache.md) — A typewriter animation over an already-complete response is not true streaming. Retained outcome: rejected.
- [Browser learning-rate default drift | PostTrainLLM](https://posttrainllm.com/experiments/browser-lr-default-drift) — [Markdown](https://posttrainllm.com/experiments/browser-lr-default-drift.md) — Reference defaults are part of correctness, not just documentation. Retained outcome: worked with caveat.
- [Browser Memory64 large-preset runs | PostTrainLLM](https://posttrainllm.com/experiments/browser-memory64-large-presets) — [Markdown](https://posttrainllm.com/experiments/browser-memory64-large-presets.md) — A reproducer must match the production host and call path before it can explain a browser-only WASM failure. Retained outcome: worked with caveat.
- [Browser speedup headline curve | PostTrainLLM](https://posttrainllm.com/experiments/browser-speedup-headline-curve) — [Markdown](https://posttrainllm.com/experiments/browser-speedup-headline-curve.md) — Qualify performance one frozen shape and machine at a time; a retained adapter identity plus alternated paired timings can convert one point without…
- [Browser WebGPU export serialization | PostTrainLLM](https://posttrainllm.com/experiments/browser-webgpu-export-serialization) — [Markdown](https://posttrainllm.com/experiments/browser-webgpu-export-serialization.md) — Long-running browser training is not product-complete until the smallest export path is smoke-tested before the long run. Retained outcome: worked with caveat.
- [Browser WebGPU training loop | PostTrainLLM](https://posttrainllm.com/experiments/browser-webgpu-training-loop) — [Markdown](https://posttrainllm.com/experiments/browser-webgpu-training-loop.md) — Hardware-qualified browser WebGPU can deliver a double-digit training-step win while preserving the frozen numerical-loss gate. Retained outcome: worked.
- [Character 2048 frontier benchmark screen (2026-08-04) | PTLLM](https://posttrainllm.com/experiments/character-2048-frontier-screen) — [Markdown](https://posttrainllm.com/experiments/character-2048-frontier-screen.md) — Validate the teacher and interface before spending compute on a specialist; a failed ruler should stop training. Retained outcome: rejected.
- [Character Chess 44.53M masked 10k pilot (2026-08-05) | PTLLM](https://posttrainllm.com/experiments/character-chess-44m-masked-10k-pilot) — [Markdown](https://posttrainllm.com/experiments/character-chess-44m-masked-10k-pilot.md) — Legal constrained decoding and lower move-prediction loss are not game intelligence; require full-game transfer before scaling a tiny specialist. Retained…
- [Character Chess benchmark candidate audit (2026-08-22) | PTLLM](https://posttrainllm.com/experiments/character-chess-benchmark-candidate) — [Markdown](https://posttrainllm.com/experiments/character-chess-benchmark-candidate.md) — A benchmark candidate set is not a benchmark until the frontier ceiling validates the ruler. Retained outcome: rejected.
- [Character game arena candidate report (2026-08-05) | PTLLM](https://posttrainllm.com/experiments/character-game-arena-candidate) — [Markdown](https://posttrainllm.com/experiments/character-game-arena-candidate.md) — A rating calculation is not a public rating until each underlying competition family passes its qualification contract. Retained outcome: rejected.
- [Cold-start bundle | PostTrainLLM](https://posttrainllm.com/experiments/cold-start-bundle) — [Markdown](https://posttrainllm.com/experiments/cold-start-bundle.md) — mmap and direct byte construction are good scaling levers, but large-model cold-start claims need large-model measurements. Retained outcome: worked with…
- [CPU speedup bundle | PostTrainLLM](https://posttrainllm.com/experiments/cpu-speedup-bundle) — [Markdown](https://posttrainllm.com/experiments/cpu-speedup-bundle.md) — The big win was fused accumulation inside the compiled trace; individual CPU tweaks need isolation before being credited. Retained outcome: worked with caveat.
- [Data-side sample packing and BPE-dropout | PTLLM](https://posttrainllm.com/experiments/data-perf-sample-packing-bpe-dropout) — [Markdown](https://posttrainllm.com/experiments/data-perf-sample-packing-bpe-dropout.md) — Data-pipeline regularizers need behavioral evals, not just stable loss curves, before being treated as quality improvements. Retained outcome: worked with…
- [Factory run schema | PostTrainLLM](https://posttrainllm.com/experiments/factory-run-schema) — [Markdown](https://posttrainllm.com/experiments/factory-run-schema.md) — A schema is necessary but not sufficient; publish checks must enforce complete run folders. Retained outcome: worked with caveat.
- [Qwen3-4B file-ops distilled specialist | PostTrainLLM](https://posttrainllm.com/experiments/fileops-qwen3-4b-distilled) — [Markdown](https://posttrainllm.com/experiments/fileops-qwen3-4b-distilled.md) — The artifact is useful only with routed-only positioning; it should not be sold as a general planner. Retained outcome: worked with caveat.
- [Qwen3-4B multibackend distilled variant | PostTrainLLM](https://posttrainllm.com/experiments/fileops-qwen3-4b-multibackend-distilled) — [Markdown](https://posttrainllm.com/experiments/fileops-qwen3-4b-multibackend-distilled.md) — Multi-backend distillation can over-specialize and needs breadth gates by default. Retained outcome: regressed.
- [Qwen3-4B ReST fused breadth recovery variant | PTLLM](https://posttrainllm.com/experiments/fileops-qwen3-4b-rest-fused) — [Markdown](https://posttrainllm.com/experiments/fileops-qwen3-4b-rest-fused.md) — ReST produced a real depth-and-safety specialist, not breadth recovery; routing is the win and general replacement is the failed hypothesis. Retained…
- [Gradient checkpointing Mac training | PostTrainLLM](https://posttrainllm.com/experiments/gradient-checkpointing-mac-training) — [Markdown](https://posttrainllm.com/experiments/gradient-checkpointing-mac-training.md) — Activation recompute is a scale-dependent lever; it should stay opt-in for models where memory, not step time, is the blocker. Retained outcome: worked with…
- [KV cache optimization bundle | PostTrainLLM](https://posttrainllm.com/experiments/kv-cache-optimization-bundle) — [Markdown](https://posttrainllm.com/experiments/kv-cache-optimization-bundle.md) — Reuse of prefilled KV is the user-visible win; cache storage policy changes need composition gates before becoming defaults. Retained outcome: worked with…
- [MLXFast SDPA and tied-embedding audit | PostTrainLLM](https://posttrainllm.com/experiments/mlxfast-sdpa-tied-embedding-audit) — [Markdown](https://posttrainllm.com/experiments/mlxfast-sdpa-tied-embedding-audit.md) — The suspected SDPA mask/contiguity and tied-embedding duplication issues were not present. Retained outcome: worked.
- [MoE dense-compute smoke | PostTrainLLM](https://posttrainllm.com/experiments/moe-dense-compute-smoke) — [Markdown](https://posttrainllm.com/experiments/moe-dense-compute-smoke.md) — MoE is a capacity feature until sparse gather/scatter or grouped matmul lands; it should not be marketed as a compute-saving win yet. Retained outcome:…
- [Needle 2 base public tool-selection gate (2026-08-31) | PTLLM](https://posttrainllm.com/experiments/needle2-base-public-gate) — [Markdown](https://posttrainllm.com/experiments/needle2-base-public-gate.md) — Tiny packaging and fast decode do not make a tool selector safe enough for action routing. Retained outcome: failed.
- [Needle 45M distractor-and-safety successor factorial (2026-09-03) | PTLLM](https://posttrainllm.com/experiments/needle2-successor-factorial) — [Markdown](https://posttrainllm.com/experiments/needle2-successor-factorial.md) — Tiny overfit proves wiring, not capacity; the interaction treatment recovered only two cases over plain training and could not learn destructive…
- [Needle 2 task-specific catalog ablation (2026-09-01) | PTLLM](https://posttrainllm.com/experiments/needle2-task-catalog-ablation) — [Markdown](https://posttrainllm.com/experiments/needle2-task-catalog-ablation.md) — Catalog restriction is a latency lever, not a capability or safety rescue for this base model. Retained outcome: regressed.
- [OffHours context-saturation boundary (2026-08-21) | PTLLM](https://posttrainllm.com/experiments/offhours-context-saturation) — [Markdown](https://posttrainllm.com/experiments/offhours-context-saturation.md) — The supported result is a neutral context-volume boundary, not a psychological or family-tension effect. Retained outcome: worked with caveat.
- [OffHours persistent-tension pilot (2026-08-21) | PTLLM](https://posttrainllm.com/experiments/offhours-persistent-tension) — [Markdown](https://posttrainllm.com/experiments/offhours-persistent-tension.md) — Unresolved narrative tension did not degrade measured work quality under the exercised conditions; a null result is a valid result. Retained outcome: worked…
- [OffHours Devin stress validation (2026-08-21) | PTLLM](https://posttrainllm.com/experiments/offhours-stress-validation) — [Markdown](https://posttrainllm.com/experiments/offhours-stress-validation.md) — A complete run is still inconclusive when the clean frontier calibration misses its preregistered gate. Retained outcome: inconclusive.
- [Open Kernel fused attention preparation | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-attention-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-20261023.md) — Fusion must preserve BF16 rounding, the exact RoPE table, cache slot writes and non-unit scale fallbacks. Retained outcome: worked with caveat.
- [Open Kernel wider attention key tile | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-attention-bn256-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-bn256-20261024.md) — Large full-decode kernels used 255 registers and 96 spills; wider tiles regressed. Retained outcome: regressed.
- [Open Kernel fixed-query FA3 scheduling | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-attention-fixedq-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-fixedq-20261024.md) — Below selected joint-head attention; preserve the existing FA3 fallback. Retained outcome: regressed.
- [Open Kernel split-K attention FP8 load compilation | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-fp8-load-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-fp8-load-20261023.md) — Python syntax and lint checks do not validate Triton backend types; masked FP8 loads need a compatible floating-point fill value. Retained outcome: failed.
- [Open Kernel joint-head FP8 paged attention | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-joint-heads-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-joint-heads-20261024.md) — Loading both contiguous KV heads in one CTA improves this measured decode regime. This result does not establish a general FP8 advantage: the separate-head…
- [Open Kernel shared-load verification attention | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-joint-verify-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-joint-verify-20261024.md) — 254-255 registers with no spills; shared loading did not make this mapping faster. Retained outcome: regressed.
- [Open Kernel attention merge/norm/gate fusion | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-merge-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-merge-20261024.md) — Below the selected joint-head attention; epilogue fusion alone did not win. Retained outcome: regressed.
- [Open Kernel additional full-context attention splits | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-more-splits-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-more-splits-20261024.md) — Doubling the full-context split target preserves correctness but does not establish a meaningful aggregate gain. Diagnostic profiles place 97.1 and 98.6…
- [Open Kernel native FP8 QK and PV attention | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-native-fp8-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-native-fp8-20261024.md) — Lower operand precision does not guarantee lower latency; layout conversion and dependent gathers are hypotheses until compiler or hardware evidence…
- [Open Kernel broad paged split-K attention | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-splitk-broad-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-splitk-broad-20261023.md) — A custom attention kernel can help decoding while losing verification; compare each dispatch regime with the current incumbent as well as the official…
- [Open Kernel narrow paged split-K attention | PTLLM](https://posttrainllm.com/experiments/openkernel-attention-splitk-narrow-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-attention-splitk-narrow-20261023.md) — Use measured dispatch boundaries to retain custom attention wins and avoid verification regressions; preserve live cache metadata, legal keys and FP32…
- [Open Kernel combined decoder layer | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-decoder-layer-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-decoder-layer-20261023.md) — Measure attention and MoE together while preserving the full layer router precision and stream ordering; component speedups do not multiply into a layer…
- [Open Kernel decoder layer with conservative MoE dispatch | PTLLM](https://posttrainllm.com/experiments/openkernel-decoder-layer-narrow-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-decoder-layer-narrow-20261023.md) — Use combined-layer evidence to narrow component optimizations, and rerun correctness and timing after restoring fallbacks. Retained outcome: worked with caveat.
- [Open Kernel single-token dense GEMV | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-dense-gemv-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-dense-gemv-20261023.md) — A single streaming GEMV can remove split-reduction overhead for one-token projections while retaining existing dispatch for other shapes. Retained outcome:…
- [Open Kernel large-tile dense pooler | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-dense-large-pooler-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-dense-large-pooler-20261024.md) — Larger pooler tiles did not beat selected dense. Retained outcome: regressed.
- [Open Kernel lossless single-token dense pooler | PTLLM](https://posttrainllm.com/experiments/openkernel-dense-lossless-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-dense-lossless-20261024.md) — Below selected dense despite a lossless weight/layout treatment. Retained outcome: regressed.
- [Open Kernel whole-row FP32 tiny-router reduction | PTLLM](https://posttrainllm.com/experiments/openkernel-dense-router-whole-row-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-dense-router-whole-row-20261024.md) — One whole-row FP32 reduction removes the cast launch for the one-token router and shows a substantial measured case gain. The diagnostic profile contains…
- [Open Kernel small-batch dense split-K and native tiny router | PTLLM](https://posttrainllm.com/experiments/openkernel-dense-splitk-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-dense-splitk-20261024.md) — Separate public-shape regimes before composing improvements; more splits add a reduction launch and may lose despite sufficient parallel work. Retained…
- [Open Kernel whole-row single-token dense and router reductions | PTLLM](https://posttrainllm.com/experiments/openkernel-dense-whole-row-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-dense-whole-row-20261024.md) — Whole-row FP32 reductions improve the measured long single-token projections while retaining the short O-projection path. QKV, pooler, unpooler and router…
- [Open Kernel final valid-local expert decoder composition | PTLLM](https://posttrainllm.com/experiments/openkernel-final-composition-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-final-composition-20261024.md) — Final layer is 10.01 percent faster by aggregate speedup ratio than the previous scored layer, but rank 283/749 missed the top-100 target. Component…
- [Open Kernel decoder-layer GEMV and attention composition | PTLLM](https://posttrainllm.com/experiments/openkernel-layer-compose-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-layer-compose-20261024.md) — Compose qualified components while preserving FP32 router input, stream ordering, cache semantics and normalization boundaries. Retained outcome: worked…
- [Open Kernel decoder layer with fused small-batch expert norm | PTLLM](https://posttrainllm.com/experiments/openkernel-layer-fused-norm-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-layer-fused-norm-20261023.md) — A wrapper-level fusion can carry into the full layer, but retain the layer router precision, stream ordering and measured MoE tile fallback. Retained…
- [Open Kernel joint-head attention layer integration | PTLLM](https://posttrainllm.com/experiments/openkernel-layer-joint-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-layer-joint-20261024.md) — Only 2.30 percent over the scored layer with overlapping intervals; failed the prior promotion gate. Retained outcome: inconclusive.
- [Open Kernel layer local expert-ID mapping | PTLLM](https://posttrainllm.com/experiments/openkernel-layer-local-ids-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-layer-local-ids-20261024.md) — Point estimate improved but reported interval overlaps the previous scored layer. Retained outcome: inconclusive.
- [Open Kernel layer M128 expert tile | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-layer-m128-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-layer-m128-20261024.md) — The global-density M128 tile change did not improve the aggregate. Retained outcome: regressed.
- [Open Kernel layer normalization and ID-map fusion | PTLLM](https://posttrainllm.com/experiments/openkernel-layer-norm-ids-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-layer-norm-ids-20261024.md) — Best prior layer check; not scored separately. Included in the final composition. Retained outcome: worked with caveat.
- [Open Kernel global-expert MoE block sizing | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-20261023.md) — Use the global expert count for routing density; do not confuse that heuristic with local expert IDs used by alignment. Retained outcome: worked with caveat.
- [Open Kernel grouped BF16-decoded NVFP4 experts | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-bf16-decode-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-bf16-decode-20261024.md) — 96 registers and zero spills do not explain the large regression; grouped GEMMs still dominate. Retained outcome: regressed.
- [Open Kernel single-token direct tensor-core experts | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-direct-dot-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-direct-dot-20261024.md) — The direct tensor-core path did not beat the selected Marlin implementation. Retained outcome: regressed.
- [Open Kernel single-token NVFP4 bit decoding and vector accumulation | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-direct-fast-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-direct-fast-20261024.md) — Bit decoding and one final reduction per output improve the earlier direct GEMV's 271.789-microsecond result, but still lose to Marlin. Unchanged…
- [Open Kernel single-token direct packed NVFP4 SIMT GEMV | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-direct-gemv-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-direct-gemv-20261024.md) — Removing route alignment is insufficient when serial unpack, scaling and repeated reductions dominate; register spilling or occupancy was not measured.…
- [Open Kernel fused small-batch expert norm | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-fused-norm-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-fused-norm-20261023.md) — Profile wrapper dispatch as well as kernel arithmetic: the existing fused norm kernel removed several small-batch framework launches while preserving FP32…
- [Open Kernel raw grouped NVFP4 tensor-core experts | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-grouped-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-grouped-20261024.md) — Correct tensor-core math does not establish an efficient implementation. The measured regression is inside GEMMs; unpacking cost, resource pressure and…
- [Open Kernel grouped NVFP4 experts with bit decoding | PTLLM](https://posttrainllm.com/experiments/openkernel-moe-grouped-bitdecode-20261024) — [Markdown](https://posttrainllm.com/experiments/openkernel-moe-grouped-bitdecode-20261024.md) — Cheaper bit decoding improves the earlier grouped implementation at 0.2087x but leaves every custom multi-token case slower than SGLang. The one-token…
- [Open Kernel broad split-K router | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-router-broad-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-router-broad-20261023.md) — A fused kernel needs a measured dispatch range; correctness alone does not qualify a performance change. Retained outcome: regressed.
- [Open Kernel large-token router tile | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-router-large-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-router-large-20261023.md) — Keep per-shape improvements separate from demonstrated aggregate gains. Retained outcome: inconclusive.
- [Open Kernel narrow split-K router | PostTrainLLM](https://posttrainllm.com/experiments/openkernel-router-narrow-20261023) — [Markdown](https://posttrainllm.com/experiments/openkernel-router-narrow-20261023.md) — Preserve tiny-token and large-token baseline paths, and judge tile choices using both target-shape timing and aggregate uncertainty. Retained outcome:…
- [Pace clarify-v1 Qwen3-4B LoRA | PostTrainLLM](https://posttrainllm.com/experiments/pace-clarify-v1-qwen4b-lora) — [Markdown](https://posttrainllm.com/experiments/pace-clarify-v1-qwen4b-lora.md) — Fine-tuning a strong base on a thin behavior slice requires coverage for every dimension that must not regress. Retained outcome: regressed.
- [Pace Gemma-3-12B unhappy-path drill winner | PTLLM](https://posttrainllm.com/experiments/pace-gemma-3-12b-drill-winner) — [Markdown](https://posttrainllm.com/experiments/pace-gemma-3-12b-drill-winner.md) — A stronger zero-shot base beat the local fine-tunes, but model-selection evidence can be superseded by later broader gates. Retained outcome: worked with…
- [Pace intent router v8 sealed rejection | PostTrainLLM](https://posttrainllm.com/experiments/pace-intent-router-v8-sealed) — [Markdown](https://posttrainllm.com/experiments/pace-intent-router-v8-sealed.md) — A fast classifier can learn the generator rather than the action policy. Retained outcome: rejected.
- [Pace next-step Apple FM availability blocker | PTLLM](https://posttrainllm.com/experiments/pace-next-step-apple-fm-20261005) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-apple-fm-20261005.md) — Separate tool choice, stateful stopping, input-format compatibility, and acceptance evidence. Retained outcome: blocked.
- [Pace next-step GLiNER zero-shot rejection | PTLLM](https://posttrainllm.com/experiments/pace-next-step-gliner-20261004) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-gliner-20261004.md) — Zero-shot label matching does not establish stateful next-step control. Retained outcome: rejected.
- [Pace next-step Jeff v1.2 base rejection | PostTrainLLM](https://posttrainllm.com/experiments/pace-next-step-jeff-base-20261005) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-jeff-base-20261005.md) — Jeff base latency did not preserve clarification, action choice or stopping; qualify each behavior before choosing a fast router. Retained outcome: rejected.
- [Pace next-step Jeff v1.2 tools rejection | PTLLM](https://posttrainllm.com/experiments/pace-next-step-jeff-tools-20261005) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-jeff-tools-20261005.md) — Jeff tools with the existing label contract remained weak on clarification and step control; a tool-oriented model is not a qualified policy replacement.…
- [Pace next-step Jeff documented tools-format rejection | PTLLM](https://posttrainllm.com/experiments/pace-next-step-jeff-tools-native-20261005) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-jeff-tools-native-20261005.md) — Jeff documented wording did not repair progress awareness; native format compatibility must be tested separately from DONE and task completion. Retained…
- [Pace next-step Jeff frozen synthetic smoke rejection | PTLLM](https://posttrainllm.com/experiments/pace-next-step-jeff-tools-smoke-20261005) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-jeff-tools-smoke-20261005.md) — The frozen Jeff synthetic smoke retained ambiguous and progress failures; disputed generator and teacher labels cannot establish acceptance. Retained…
- [Pace next-step Jev teacher development baseline | PTLLM](https://posttrainllm.com/experiments/pace-next-step-jev-20261004) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-jev-20261004.md) — Use the teacher as offline comparison evidence; runtime integration requires an on-device candidate. Retained outcome: worked with caveat.
- [Pace next-step Qwen development baseline | PTLLM](https://posttrainllm.com/experiments/pace-next-step-qwen-20261004) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-qwen-20261004.md) — Compare identical simulated loops and argument extractors rather than model-card scores. Retained outcome: worked with caveat.
- [Pace next-step Qwen reproduced development baseline | PTLLM](https://posttrainllm.com/experiments/pace-next-step-qwen-20261005) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-qwen-20261005.md) — Separate tool choice, stateful stopping, input-format compatibility, and acceptance evidence. Retained outcome: worked with caveat.
- [Pace frozen generated policy Qwen comparison | PTLLM](https://posttrainllm.com/experiments/pace-next-step-qwen-policy-20261008) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-qwen-policy-20261008.md) — Preserve raw failures, generated provenance and unknown calibration/hardware measures; a fast or rules-perfect result does not qualify a planner…
- [Pace next-step ModernBERT student v1 | PostTrainLLM](https://posttrainllm.com/experiments/pace-next-step-student-v1-20261004) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-student-v1-20261004.md) — Student v1 left single-target clicks weak; label policy and independent evaluation must precede further teacher distillation. Retained outcome: inconclusive.
- [Pace next-step ModernBERT student v2 | PostTrainLLM](https://posttrainllm.com/experiments/pace-next-step-student-v2-20261004) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-student-v2-20261004.md) — Agreement filtering lowered student v2 full pass to 53.7%; agreement alone does not repair a disputed action policy. Retained outcome: inconclusive.
- [Pace next-step ModernBERT student v3 | PostTrainLLM](https://posttrainllm.com/experiments/pace-next-step-student-v3-20261004) — [Markdown](https://posttrainllm.com/experiments/pace-next-step-student-v3-20261004.md) — Goal-phrased training improved student v3 clicks but regressed RESPOND; report per-class gates alongside aggregate full pass. Retained outcome: inconclusive.
- [Pace planner v1-v4 early versions | PostTrainLLM](https://posttrainllm.com/experiments/pace-planner-v1-v4-early) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v1-v4-early.md) — An unmeasured version is indistinguishable from a version that never ran. Four consecutive attempts left no score behind. Retained outcome: inconclusive.
- [Pace planner v10 parameterized actions | PostTrainLLM](https://posttrainllm.com/experiments/pace-planner-v10) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v10.md) — v10 repeated v6's move of changing the output representation, and because it was never measured the v6 regression could not be checked for recurrence. Its…
- [Pace planner v11 ship-gate run | PostTrainLLM](https://posttrainllm.com/experiments/pace-planner-v11) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v11.md) — Train loss 0.001 alongside 0% on out-of-scope and ambiguous inputs is the capacity ceiling stated plainly. The one-shot fixed gate was what finally ended…
- [Pace planner v5 hand-crafted gold + teacher paraphrase | PTLLM](https://posttrainllm.com/experiments/pace-planner-v5) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v5.md) — A high score on a fixture a deterministic endpoint can also ace is a measurement of the harness, not the model. Retained outcome: worked with caveat.
- [Pace planner v6 label-based output | PostTrainLLM](https://posttrainllm.com/experiments/pace-planner-v6) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v6.md) — Changing the output representation is a capability change, not a formatting change. Retained outcome: regressed.
- [Pace planner v6.1 four-adapter repair sweep | PTLLM](https://posttrainllm.com/experiments/pace-planner-v6-1) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v6-1.md) — A four-way score collapse traced to prompt mismatch between training and serving. Pin the prompt, schema, and grammar triple across train, serve, and eval…
- [Pace planner v7 tools-in-prompt | PostTrainLLM](https://posttrainllm.com/experiments/pace-planner-v7) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v7.md) — v7 is a numbering gap, not a training version. The eleven-version count includes a version that was never trained. Retained outcome: superseded.
- [Pace planner v8 augmented corpus | PostTrainLLM](https://posttrainllm.com/experiments/pace-planner-v8) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v8.md) — The 73% phantom. A score is only a score if the prompt, schema, and grammar it was measured under are recorded with it. Retained outcome: inconclusive.
- [Pace planner v9 compose | PostTrainLLM](https://posttrainllm.com/experiments/pace-planner-v9) — [Markdown](https://posttrainllm.com/experiments/pace-planner-v9.md) — Winning the happy-path fixture says nothing about ambiguous, out-of-scope, or destructive inputs, which is where the product actually needed the planner.…
- [Pace Qwen3.5 challenger round | PostTrainLLM](https://posttrainllm.com/experiments/pace-qwen35-challenger-round) — [Markdown](https://posttrainllm.com/experiments/pace-qwen35-challenger-round.md) — Single-dimension wins do not justify a planner swap when another required dimension regresses. Retained outcome: worked with caveat.
- [Pace reasoning-model unhappy-path drill | PostTrainLLM](https://posttrainllm.com/experiments/pace-thinking-models-unhappy-path) — [Markdown](https://posttrainllm.com/experiments/pace-thinking-models-unhappy-path.md) — Reasoning models need adequate max tokens and a stop-on-think strategy before interpreting planner scores. Retained outcome: inconclusive.
- [Pace two-stage ambiguity detector shim v2 | PTLLM](https://posttrainllm.com/experiments/pace-two-stage-ambiguity-shim-v2) — [Markdown](https://posttrainllm.com/experiments/pace-two-stage-ambiguity-shim-v2.md) — Rule wrappers can expose the missing behavior, but they still need the same regression gates as model updates. Retained outcome: worked with caveat.
- [Parakeet WGSL browser-ASR paired validation (2026-09-02) | PTLLM](https://posttrainllm.com/experiments/parakeet-wgsl-browser-asr-smoke) — [Markdown](https://posttrainllm.com/experiments/parakeet-wgsl-browser-asr-smoke.md) — Browser WebGPU ASR can beat a native large-model incumbent on both bounded WER and latency while still failing an absolute throughput target; short-clip…
- [Public artifact registry | PostTrainLLM](https://posttrainllm.com/experiments/public-artifact-registry) — [Markdown](https://posttrainllm.com/experiments/public-artifact-registry.md) — Public artifacts and blockers are first-class. Retained outcome: worked.
- [Speculative-decoding heads Medusa/EAGLE smoke | PTLLM](https://posttrainllm.com/experiments/speculative-heads-medusa-eagle-smoke) — [Markdown](https://posttrainllm.com/experiments/speculative-heads-medusa-eagle-smoke.md) — Speculative heads are a valid local speed path, but the production recipe is data/training heavy rather than a quick sidecar smoke. Retained outcome: worked…
- [Blended public+synthetic SFT | PostTrainLLM](https://posttrainllm.com/experiments/sql-blended-public-synthetic-v1) — [Markdown](https://posttrainllm.com/experiments/sql-blended-public-synthetic-v1.md) — Mixing domains is not automatically regularization; it can reduce both capabilities. Retained outcome: failed.
- [SQL candidate selection | PostTrainLLM](https://posttrainllm.com/experiments/sql-candidate-selection) — [Markdown](https://posttrainllm.com/experiments/sql-candidate-selection.md) — Candidate selection remains a plausible future method, but an unfrozen recipe is not unfinished evidence for this project. Retained outcome: rejected.
- [Controlled SQL LoRA rank sweep | PostTrainLLM](https://posttrainllm.com/experiments/sql-controlled-lora-rank-sweep) — [Markdown](https://posttrainllm.com/experiments/sql-controlled-lora-rank-sweep.md) — Hyperparameter sweeps are valid only inside one frozen experiment contract; they are not project-completion chores. Retained outcome: rejected.
- [Expanded synthetic SQL SFT | PostTrainLLM](https://posttrainllm.com/experiments/sql-expanded-synthetic-sft-r4) — [Markdown](https://posttrainllm.com/experiments/sql-expanded-synthetic-sft-r4.md) — The synthetic SQL adapter is the current synthetic incumbent, but joins remain weaker than single-table rows. Retained outcome: worked.
- [SQL factory run renderer | PostTrainLLM](https://posttrainllm.com/experiments/sql-factory-run-renderer) — [Markdown](https://posttrainllm.com/experiments/sql-factory-run-renderer.md) — Rendered reports made SQL publishable, but the factory still needs command-level automation. Retained outcome: worked with caveat.
- [Hygiene DPO higher-pressure retry (2026-07-11) | PTLLM](https://posttrainllm.com/experiments/sql-hygiene-dpo-higher-pressure) — [Markdown](https://posttrainllm.com/experiments/sql-hygiene-dpo-higher-pressure.md) — Composed DPO cannot fix output-format hygiene here Retained outcome: failed.
- [Hygiene DPO reference-anchored retry (2026-07-11) | PTLLM](https://posttrainllm.com/experiments/sql-hygiene-dpo-refanchored) — [Markdown](https://posttrainllm.com/experiments/sql-hygiene-dpo-refanchored.md) — Reference-anchored DPO is the correct fix for the SimPO collapse Retained outcome: failed.
- [Hygiene SimPO/DPO | PostTrainLLM](https://posttrainllm.com/experiments/sql-hygiene-simpo-dpo) — [Markdown](https://posttrainllm.com/experiments/sql-hygiene-simpo-dpo.md) — Preference tuning needs reference anchoring, smaller updates, and composed-eval checks. Retained outcome: failed.
- [Offline rollout / OAPL-style SQL update | PostTrainLLM](https://posttrainllm.com/experiments/sql-offline-rollout-oapl) — [Markdown](https://posttrainllm.com/experiments/sql-offline-rollout-oapl.md) — Offline policy updates are downstream of reward validation; skipping the prerequisite is not a useful experiment. Retained outcome: rejected.
- [Public b-mc2 v1 | PostTrainLLM](https://posttrainllm.com/experiments/sql-public-bmc2-v1) — [Markdown](https://posttrainllm.com/experiments/sql-public-bmc2-v1.md) — Public benchmark adaptation needs more coverage before tuning knobs matter. Retained outcome: failed.
- [Public b-mc2 v2 | PostTrainLLM](https://posttrainllm.com/experiments/sql-public-bmc2-v2) — [Markdown](https://posttrainllm.com/experiments/sql-public-bmc2-v2.md) — Do not change data volume, rank, and training length in one uncontrolled jump. Retained outcome: failed.
- [Public b-mc2 v3 | PostTrainLLM](https://posttrainllm.com/experiments/sql-public-bmc2-v3) — [Markdown](https://posttrainllm.com/experiments/sql-public-bmc2-v3.md) — The public recipe was directionally better, but not enough for a win. Retained outcome: failed.
- [Public b-mc2 v4 | PostTrainLLM](https://posttrainllm.com/experiments/sql-public-bmc2-v4) — [Markdown](https://posttrainllm.com/experiments/sql-public-bmc2-v4.md) — A narrow public exact-match win can still damage the incumbent specialist behavior. Retained outcome: worked with caveat.
- [Public v4 synthetic regression check | PostTrainLLM](https://posttrainllm.com/experiments/sql-public-v4-synthetic-regression) — [Markdown](https://posttrainllm.com/experiments/sql-public-v4-synthetic-regression.md) — Every public benchmark gain needs a breadth/regression gate before packaging. Retained outcome: regressed.
- [Routed public + synthetic adapters | PostTrainLLM](https://posttrainllm.com/experiments/sql-routed-v1) — [Markdown](https://posttrainllm.com/experiments/sql-routed-v1.md) — Routing preserves both measured slices, but artifact status still depends on public execution and performance evidence. Retained outcome: worked with caveat.
- [Static multi-LoRA composition | PostTrainLLM](https://posttrainllm.com/experiments/sql-static-multilora-composition) — [Markdown](https://posttrainllm.com/experiments/sql-static-multilora-composition.md) — Static composition cannot solve domain conflict when the right behavior depends on input type. Retained outcome: failed.
- [Toy SQL SFT | PostTrainLLM](https://posttrainllm.com/experiments/sql-toy-sft-r4) — [Markdown](https://posttrainllm.com/experiments/sql-toy-sft-r4.md) — The SFT loop mechanics work, but eval splits must be frozen before claiming quality. Retained outcome: worked with caveat.
- [StreamingLLM + KIVI cache compression | PostTrainLLM](https://posttrainllm.com/experiments/streamingllm-kivi-cache-compression) — [Markdown](https://posttrainllm.com/experiments/streamingllm-kivi-cache-compression.md) — KV compression can preserve greedy decode locally, but storage format and quality gates matter before making long-context claims. Retained outcome: worked…
- [Technique registry | PostTrainLLM](https://posttrainllm.com/experiments/technique-registry) — [Markdown](https://posttrainllm.com/experiments/technique-registry.md) — Method names are not enough; recipes need failure mode, data, eval, slice gate, and stop rule. Retained outcome: worked with caveat.
- [TrainLoop-style tooling | PostTrainLLM](https://posttrainllm.com/experiments/trainloop-style-tooling) — [Markdown](https://posttrainllm.com/experiments/trainloop-style-tooling.md) — Discipline is valuable only when enforced in every reported run. Retained outcome: worked with caveat.
- [VibeThinker 3B agentic distilled variant | PTLLM](https://posttrainllm.com/experiments/vibethinker-3b-agentic-distilled) — [Markdown](https://posttrainllm.com/experiments/vibethinker-3b-agentic-distilled.md) — Preserved weights are not a ship decision. Retained outcome: inconclusive.
- [WASM backward_dA blocking | PostTrainLLM](https://posttrainllm.com/experiments/wasm-backward-da-blocking) — [Markdown](https://posttrainllm.com/experiments/wasm-backward-da-blocking.md) — Training speed follows the backward-pass bottleneck; forward-only wins are capped until dA and dB are addressed. Retained outcome: worked with caveat.
- [WASM backward-scratch reuse | PostTrainLLM](https://posttrainllm.com/experiments/wasm-backward-scratch-reuse) — [Markdown](https://posttrainllm.com/experiments/wasm-backward-scratch-reuse.md) — Hot-path hygiene is not the same as a performance win; measure the real step time before claiming speedup. Retained outcome: worked with caveat.
- [WASM register/cache-blocked matmul | PostTrainLLM](https://posttrainllm.com/experiments/wasm-register-cache-blocked-matmul) — [Markdown](https://posttrainllm.com/experiments/wasm-register-cache-blocked-matmul.md) — Optimize and report on the real workload, not only on isolated single-kernel microbenches. Retained outcome: worked with caveat.
- [WASM SIMD training speedup | PostTrainLLM](https://posttrainllm.com/experiments/wasm-simd-training-speedup) — [Markdown](https://posttrainllm.com/experiments/wasm-simd-training-speedup.md) — Emscripten/LLVM autovectorization was the right browser-training speed lever before hand-written kernels. Retained outcome: worked.
- [YOCO cache-halving smoke | PostTrainLLM](https://posttrainllm.com/experiments/yoco-cache-halving-smoke) — [Markdown](https://posttrainllm.com/experiments/yoco-cache-halving-smoke.md) — YOCO is a long-context memory lever, not an immediate short-context speedup. Retained outcome: worked with caveat.

## Reproducible recipe contracts

- [Constrained generation | PostTrainLLM](https://posttrainllm.com/recipes/constrained-generation) — [Markdown](https://posttrainllm.com/recipes/constrained-generation.md) — Structured tool calls, SQL, and schema-bound outputs. A capable model emits invalid syntax or forbidden output shapes.
- [Specialist distillation | PostTrainLLM](https://posttrainllm.com/recipes/distillation) — [Markdown](https://posttrainllm.com/recipes/distillation.md) — Compress a frontier or larger-model capability into a routed Mac-local specialist.. A small base lacks a narrow behavior that a qualified teacher can…
- [Evolution strategies | PostTrainLLM](https://posttrainllm.com/recipes/evolution-strategies) — [Markdown](https://posttrainllm.com/recipes/evolution-strategies.md) — Gradient-free optimization when rewards are evaluable but differentiation is unavailable.. The objective cannot provide useful gradients or the model/policy…
- [GaLore and training stability | PostTrainLLM](https://posttrainllm.com/recipes/galore-stability) — [Markdown](https://posttrainllm.com/recipes/galore-stability.md) — Reduce optimizer memory or stabilize deep training.. Optimizer state or unstable gradients block an otherwise valid training recipe.
- [Interpretability probes | PostTrainLLM](https://posttrainllm.com/recipes/interpretability) — [Markdown](https://posttrainllm.com/recipes/interpretability.md) — Diagnose a measured model failure without treating a visualization as causal proof.. Aggregate scores cannot explain which layers, tokens, or…
- [LoRA, DoRA, and adapter geometry | PostTrainLLM](https://posttrainllm.com/recipes/lora) — [Markdown](https://posttrainllm.com/recipes/lora.md) — Adapt a frozen Mac-local base to one narrow behavior.. The base misses a bounded task and full fine-tuning is unnecessary or too costly.
- [Sparse mixture of experts | PostTrainLLM](https://posttrainllm.com/recipes/moe) — [Markdown](https://posttrainllm.com/recipes/moe.md) — Increase conditional capacity while keeping active compute bounded.. A dense model lacks capacity and a true sparse execution path is available.
- [Multi-token prediction | PostTrainLLM](https://posttrainllm.com/recipes/mtp) — [Markdown](https://posttrainllm.com/recipes/mtp.md) — Improve training signal or support speculative multi-token heads.. Single-token supervision or decode throughput is the measured bottleneck.
- [Needle 45M successor factorial | PostTrainLLM](https://posttrainllm.com/recipes/needle2-evaluation) — [Markdown](https://posttrainllm.com/recipes/needle2-evaluation.md) — Test whether distractor-rich selection data and explicit refusal/confirmation examples can make a 45M tool selector safe and more accurate than stock.. Tiny…
- [Optimizer and schedule selection | PostTrainLLM](https://posttrainllm.com/recipes/optimizers) — [Markdown](https://posttrainllm.com/recipes/optimizers.md) — Make a correct training recipe converge stably within Mac memory and time bounds.. Loss spikes, under-updating, or optimizer memory blocks a frozen recipe.
- [Parakeet WGSL browser ASR | PostTrainLLM](https://posttrainllm.com/recipes/parakeet-browser-asr) — [Markdown](https://posttrainllm.com/recipes/parakeet-browser-asr.md) — Test whether long-form ASR can run locally in Chrome through WebGPU.. Browser audio ingestion may be too slow, too large, nondeterministic, or…
- [Parameter-efficient fine-tuning variants | PTLLM](https://posttrainllm.com/recipes/peft-variants) — [Markdown](https://posttrainllm.com/recipes/peft-variants.md) — Trade adapter capacity, initialization, and storage for one narrow fine-tuning target.. Plain LoRA has a measured geometry, capacity, quantization, or…
- [Precision and numerical gates | PostTrainLLM](https://posttrainllm.com/recipes/precision) — [Markdown](https://posttrainllm.com/recipes/precision.md) — Reduce memory or improve throughput without invalidating model math.. Precision choice causes overflow, drift, excess memory, or slow execution.
- [Unstructured and structured pruning | PostTrainLLM](https://posttrainllm.com/recipes/pruning) — [Markdown](https://posttrainllm.com/recipes/pruning.md) — Reduce stored or executed model structure after capability is proven.. Artifact size or dense topology is the measured deployment bottleneck.
- [Quantization and storage formats | PostTrainLLM](https://posttrainllm.com/recipes/quantization) — [Markdown](https://posttrainllm.com/recipes/quantization.md) — Reduce model download, RAM, or execution cost on Mac and browser runtimes.. Full precision exceeds the target's distribution or memory budget.
- [Speculative heads and decoding | PostTrainLLM](https://posttrainllm.com/recipes/speculative-heads) — [Markdown](https://posttrainllm.com/recipes/speculative-heads.md) — Increase local decode throughput after model quality is accepted.. Autoregressive decode latency is the measured user-visible bottleneck.
- [SQL specialist lineage | PostTrainLLM](https://posttrainllm.com/recipes/sql-lineage) — [Markdown](https://posttrainllm.com/recipes/sql-lineage.md) — Generate or select executable SQL on frozen schemas.. Small-model SQL gains fail to generalize across public and synthetic domains or emit prose wrappers.
- [StreamingLLM and KIVI cache compression | PostTrainLLM](https://posttrainllm.com/recipes/streaming-kivi) — [Markdown](https://posttrainllm.com/recipes/streaming-kivi.md) — Bound or compress KV memory for long local contexts.. KV cache growth, not model weights, is the measured memory blocker.

## Learning paths

- [Architecture and kernel execution | PostTrainLLM](https://posttrainllm.com/learn/paths/architecture-and-kernels) — [Markdown](https://posttrainllm.com/learn/paths/architecture-and-kernels.md) — Match one optimized attention or matmul path against the readable reference and identify the real bottleneck with a measured curve. Mastery is demonstrated…
- [Browser, WASM, WebGPU, MLX, and Apple runtime | PTLLM](https://posttrainllm.com/learn/paths/browser-and-mac-runtime) — [Markdown](https://posttrainllm.com/learn/paths/browser-and-mac-runtime.md) — Compare the historical unqualified WebGPU curve and its missing receipt, Memory64 limit, Apple action-grounding result, and Parakeet browser receipt as four…
- [Evaluation, evidence, and the specialist factory | PTLLM](https://posttrainllm.com/learn/paths/evaluation-and-factory) — [Markdown](https://posttrainllm.com/learn/paths/evaluation-and-factory.md) — Audit Needle, Chess, and 2048 and explain how the ruler or safety gate changed the decision before integration or training. Mastery is demonstrated when the…
- [Foundations to transformer | PostTrainLLM](https://posttrainllm.com/learn/paths/foundations) — [Markdown](https://posttrainllm.com/learn/paths/foundations.md) — Complete the toy calculation or implementation in Modules 1-7 and explain its tensor shapes aloud. Mastery is demonstrated when the learner can trace text…
- [Interpretability with causal discipline | PostTrainLLM](https://posttrainllm.com/learn/paths/interpretability) — [Markdown](https://posttrainllm.com/learn/paths/interpretability.md) — Take one behavioral failure and write a probe hypothesis, negative control, intervention, and decision-changing prediction. Mastery is demonstrated when the…
- [SFT, adapters, preference tuning, and distillation | PTLLM](https://posttrainllm.com/learn/paths/post-training) — [Markdown](https://posttrainllm.com/learn/paths/post-training.md) — Reconstruct the SQL and file-ops lineages, including the variable changed, target gain, regression, and decision at each step. Mastery is demonstrated when…
- [Quantization, pruning, and artifact packaging | PTLLM](https://posttrainllm.com/learn/paths/quantization-and-packaging) — [Markdown](https://posttrainllm.com/learn/paths/quantization-and-packaging.md) — Audit one int4 or pruned artifact from source weights through file size, load parity, quality, RAM, and runtime measurement. Mastery is demonstrated when…
- [Inference, serving, tools, and long context | PTLLM](https://posttrainllm.com/learn/paths/runtime-and-agents) — [Markdown](https://posttrainllm.com/learn/paths/runtime-and-agents.md) — Calculate KV memory, compare cold versus reused-prefill latency, and trace one tool request through model output, validation, and execution policy. Mastery…
- [Training mechanics and stability | PostTrainLLM](https://posttrainllm.com/learn/paths/training-mechanics) — [Markdown](https://posttrainllm.com/learn/paths/training-mechanics.md) — Inspect one successful tiny-overfit and one failed ordinary run, then diagnose data, optimization, capacity, and evaluation separately. Mastery is…

## Buildable learning artifacts

- [Parakeet browser ASR receipt | PostTrainLLM](https://posttrainllm.com/learn/artifacts/browser-asr-case) — [Markdown](https://posttrainllm.com/learn/artifacts/browser-asr-case.md) — A paired browser/native ASR lab: Parakeet wins bounded WER and latency while a frozen 50x short-clip throughput gate catches fixed dispatch cost.
- [Browser-trained TinyGPT | PostTrainLLM](https://posttrainllm.com/learn/artifacts/browser-tinygpt) — [Markdown](https://posttrainllm.com/learn/artifacts/browser-tinygpt.md) — The same learning loop made visible through loss, sampling, presets, workers, WASM, and WebGPU-capable browser surfaces.
- [Byte-level TinyGPT | PostTrainLLM](https://posttrainllm.com/learn/artifacts/byte-tinygpt) — [Markdown](https://posttrainllm.com/learn/artifacts/byte-tinygpt.md) — A readable ~0.8M-parameter transformer whose byte tokenizer, architecture, loss, and checkpoint format are all inspectable.
- [Causal probe dossier | PostTrainLLM](https://posttrainllm.com/learn/artifacts/causal-probe-dossier) — [Markdown](https://posttrainllm.com/learn/artifacts/causal-probe-dossier.md) — A logit lens, linear probe, activation patch, causal trace, or sparse autoencoder investigation with negative controls and an intervention claim.
- [Distilled tool-calling specialist | PostTrainLLM](https://posttrainllm.com/learn/artifacts/distilled-tool-caller) — [Markdown](https://posttrainllm.com/learn/artifacts/distilled-tool-caller.md) — A real lineage where narrow file-operation depth reached 100% while breadth regressed, proving why routing is part of the artifact.
- [Specialist factory report card | PostTrainLLM](https://posttrainllm.com/learn/artifacts/factory-report-card) — [Markdown](https://posttrainllm.com/learn/artifacts/factory-report-card.md) — A run directory that binds target, data, candidate, frozen evaluation, packaging evidence, resource measurements, and a terminal ship/retry/reject decision.
- [Kernel parity and speed receipt | PostTrainLLM](https://posttrainllm.com/learn/artifacts/kernel-parity-receipt) — [Markdown](https://posttrainllm.com/learn/artifacts/kernel-parity-receipt.md) — A numerical and performance comparison that keeps reference math, optimized kernel, dispatch geometry, and end-to-end effect separate.
- [Local tool-using agent | PostTrainLLM](https://posttrainllm.com/learn/artifacts/local-tool-agent) — [Markdown](https://posttrainllm.com/learn/artifacts/local-tool-agent.md) — An OpenAI-compatible local model server plus a multi-turn tool loop with schema validation and explicit execution policy.
- [LoRA specialist adapter | PostTrainLLM](https://posttrainllm.com/learn/artifacts/lora-specialist) — [Markdown](https://posttrainllm.com/learn/artifacts/lora-specialist.md) — A narrow adapter built from a frozen base, traceable data, explicit LoRA geometry, and a held-out task gate.
- [Mac runtime boundary map | PostTrainLLM](https://posttrainllm.com/learn/artifacts/mac-runtime-boundary-map) — [Markdown](https://posttrainllm.com/learn/artifacts/mac-runtime-boundary-map.md) — A comparison of owned MLX weights, Core ML deployment, Apple Foundation Models, and browser runtimes grounded in measured capability limits.
- [Validated MLX specialist package | PostTrainLLM](https://posttrainllm.com/learn/artifacts/mlx-specialist-package) — [Markdown](https://posttrainllm.com/learn/artifacts/mlx-specialist-package.md) — A model or adapter exported with identity, tensors, tokenizer, prompt contract, and evidence that the consumer can load it correctly.
- [Needle 2 selection ruler | PostTrainLLM](https://posttrainllm.com/learn/artifacts/needle-selection-ruler) — [Markdown](https://posttrainllm.com/learn/artifacts/needle-selection-ruler.md) — A complete tiny-model journey from catalog ablation through a 2x2 distractor-and-safety training factorial, ending at a measured 45M capacity boundary.
- [Tiny-overfit correctness receipt | PostTrainLLM](https://posttrainllm.com/learn/artifacts/tiny-overfit-receipt) — [Markdown](https://posttrainllm.com/learn/artifacts/tiny-overfit-receipt.md) — A minimal proof that data loading, forward pass, backpropagation, optimizer state, sampling, and checkpoint reload agree.

## Fine-tune report cards

- [Fine-Tune Report Card — sql-routed-specialist-poc — routed-sft-lora](https://posttrainllm.com/report-cards/qwen06-sql-routed-v1) — [Markdown](https://posttrainllm.com/report-cards/qwen06-sql-routed-v1.md) — Retry — eval: The routed setup is the current best SQL candidate and passes the current public exact plus synthetic execution gates, but it is not a shipped…
- [Fine-Tune Report Card — Qwen3-4B File-Ops Distilled](https://posttrainllm.com/report-cards/qwen3-4b-file-ops-distilled) — [Markdown](https://posttrainllm.com/report-cards/qwen3-4b-file-ops-distilled.md) — Shipped: ship only as a routed file-ops specialist; do not use as the general planner
- [Fine-Tune Report Card — Qwen3-4B ReST Fused](https://posttrainllm.com/report-cards/qwen3-4b-rest-fused) — [Markdown](https://posttrainllm.com/report-cards/qwen3-4b-rest-fused.md) — Shipped: retain only as a routed file-operations specialist; reject as a general successor and do not use as the Pace default planner

## Documentation

- [PostTrainLLM docs technical documentation](https://posttrainllm.com/docs/) — [Markdown](https://posttrainllm.com/docs/index.md) — Mac-local LLM specialist factory — training, inference, evals, systems notes, and learning paths.
- [Citations | PostTrainLLM reference](https://posttrainllm.com/docs/CITATIONS/) — [Markdown](https://posttrainllm.com/docs/CITATIONS.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [GRPO-Clarify v1 — RL on the ambig dimension](https://posttrainllm.com/docs/GRPO_CLARIFY/) — [Markdown](https://posttrainllm.com/docs/GRPO_CLARIFY.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Doc map — where things moved | PostTrainLLM reference](https://posttrainllm.com/docs/MAP/) — [Markdown](https://posttrainllm.com/docs/MAP.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM Closure and Fresh-Experiment Gate](https://posttrainllm.com/docs/NEXT/) — [Markdown](https://posttrainllm.com/docs/NEXT.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM — master plan (shipped / skipped / TODO)](https://posttrainllm.com/docs/PLAN/) — [Markdown](https://posttrainllm.com/docs/PLAN.md) — Single source of truth for what's shipped, skipped, and still to build. Consolidated from docs/roadmap/*, docs/progress.md, docs/backlog.md,…
- [Project Recommendation Context](https://posttrainllm.com/docs/PROJECT_RECOMMENDATION_CONTEXT/) — [Markdown](https://posttrainllm.com/docs/PROJECT_RECOMMENDATION_CONTEXT.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM Docs technical documentation](https://posttrainllm.com/docs/README/) — [Markdown](https://posttrainllm.com/docs/README.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Agent runtime | PostTrainLLM reference](https://posttrainllm.com/docs/agent_runtime/) — [Markdown](https://posttrainllm.com/docs/agent_runtime.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [How it works: the specialist factory end-to-end](https://posttrainllm.com/docs/architecture/how-it-works/) — [Markdown](https://posttrainllm.com/docs/architecture/how-it-works.md) — A learning-first walkthrough of posttrainllm's Mac-local specialist factory — the target → data → post-training → eval → package → report loop, the native…
- [I built a GPT-2 in the browser, then made it 2.6×–12.1×…](https://posttrainllm.com/docs/archive/BLOG/) — [Markdown](https://posttrainllm.com/docs/archive/BLOG.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session handoff — pick this up cleanly](https://posttrainllm.com/docs/archive/HANDOFF/) — [Markdown](https://posttrainllm.com/docs/archive/HANDOFF.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Milestones | PostTrainLLM reference](https://posttrainllm.com/docs/archive/MILESTONES/) — [Markdown](https://posttrainllm.com/docs/archive/MILESTONES.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [While you slept | PostTrainLLM reference](https://posttrainllm.com/docs/archive/WHILE_YOU_SLEPT/) — [Markdown](https://posttrainllm.com/docs/archive/WHILE_YOU_SLEPT.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Archived AGENTS.md history (moved 2026-10-08)](https://posttrainllm.com/docs/archive/agents-history/) — [Markdown](https://posttrainllm.com/docs/archive/agents-history.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Annotated transcript | PostTrainLLM reference](https://posttrainllm.com/docs/archive/annotated_transcript/) — [Markdown](https://posttrainllm.com/docs/archive/annotated_transcript.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Evaluation and safety | PostTrainLLM reference](https://posttrainllm.com/docs/archive/evaluation/) — [Markdown](https://posttrainllm.com/docs/archive/evaluation.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Learning roadmap | PostTrainLLM reference](https://posttrainllm.com/docs/archive/learning_roadmap/) — [Markdown](https://posttrainllm.com/docs/archive/learning_roadmap.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Lessons from building posttrainllm](https://posttrainllm.com/docs/archive/lessons/) — [Markdown](https://posttrainllm.com/docs/archive/lessons.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Parked: the four "multi-model" directions](https://posttrainllm.com/docs/archive/parked_multi_model/) — [Markdown](https://posttrainllm.com/docs/archive/parked_multi_model.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Phase 9 + 10 — status and follow-up design](https://posttrainllm.com/docs/archive/phase_9_10_status/) — [Markdown](https://posttrainllm.com/docs/archive/phase_9_10_status.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Roadmap — recent research (2024-2026 highlights)](https://posttrainllm.com/docs/archive/recent_research/) — [Markdown](https://posttrainllm.com/docs/archive/recent_research.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Shared vs. native — how the browser and macOS paths relate](https://posttrainllm.com/docs/archive/shared_vs_native/) — [Markdown](https://posttrainllm.com/docs/archive/shared_vs_native.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Project status — 2026 update | PostTrainLLM reference](https://posttrainllm.com/docs/archive/status/) — [Markdown](https://posttrainllm.com/docs/archive/status.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Watch the model think | PostTrainLLM reference](https://posttrainllm.com/docs/archive/watch_the_model_think/) — [Markdown](https://posttrainllm.com/docs/archive/watch_the_model_think.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [GitHub Corpus Parity Decision | PostTrainLLM reference](https://posttrainllm.com/docs/artifacts/github-corpus-parity/) — [Markdown](https://posttrainllm.com/docs/artifacts/github-corpus-parity.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Async tool-call dispatch — investigated, skipped](https://posttrainllm.com/docs/async_tool_dispatch/) — [Markdown](https://posttrainllm.com/docs/async_tool_dispatch.md) — Why async tool dispatch in AgentLoop doesn't justify the work at current bottleneck shape. Logged so the decision doesn't get re-litigated.
- [Attempt Ledger | PostTrainLLM reference](https://posttrainllm.com/docs/attempt-ledger/) — [Markdown](https://posttrainllm.com/docs/attempt-ledger.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Audit 2026 — what we tried, what worked, what we flagged](https://posttrainllm.com/docs/audits/audit_2026/) — [Markdown](https://posttrainllm.com/docs/audits/audit_2026.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Docs Quality Audit | PostTrainLLM reference](https://posttrainllm.com/docs/audits/docs-quality-audit/) — [Markdown](https://posttrainllm.com/docs/audits/docs-quality-audit.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Exactness Completion Audit | PostTrainLLM reference](https://posttrainllm.com/docs/audits/exactness-completion-audit/) — [Markdown](https://posttrainllm.com/docs/audits/exactness-completion-audit.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Feature audit (2026-05-31) — moved](https://posttrainllm.com/docs/audits/feature_audit_2026_05_31/) — [Markdown](https://posttrainllm.com/docs/audits/feature_audit_2026_05_31.md) — Verification results absorbed into docs/PLAN.md §1. Kept here only as the audit's historical anchor.
- [History Coverage Audit | PostTrainLLM reference](https://posttrainllm.com/docs/audits/history-coverage-audit/) — [Markdown](https://posttrainllm.com/docs/audits/history-coverage-audit.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Test Coverage Goals — posttrainllm](https://posttrainllm.com/docs/audits/test-coverage/) — [Markdown](https://posttrainllm.com/docs/audits/test-coverage.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Phase 1-10 validation + end-to-end workflows](https://posttrainllm.com/docs/audits/validation_report/) — [Markdown](https://posttrainllm.com/docs/audits/validation_report.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [ROI-ordered backlog — moved | PostTrainLLM reference](https://posttrainllm.com/docs/backlog/) — [Markdown](https://posttrainllm.com/docs/backlog.md) — Consolidated into docs/PLAN.md §3 (TODO). Tier A/B/C/D ordering preserved.
- [BPE-aware leaderboard scoring | PostTrainLLM reference](https://posttrainllm.com/docs/bpe_browser_scoring/) — [Markdown](https://posttrainllm.com/docs/bpe_browser_scoring.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Browser notes — WASM, Workers, OPFS, WebGPU](https://posttrainllm.com/docs/browser_notes/) — [Markdown](https://posttrainllm.com/docs/browser_notes.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Capability matrix — modalities × actions](https://posttrainllm.com/docs/capability_matrix/) — [Markdown](https://posttrainllm.com/docs/capability_matrix.md) — Exhaustive map of input modalities posttrainllm supports and what operations (train, distill, tune, quantize, etc.) are available for each. Honest status per…
- [CLI lab reference | PostTrainLLM reference](https://posttrainllm.com/docs/cli-reference/) — [Markdown](https://posttrainllm.com/docs/cli-reference.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Dataset inventory — what's available, sizes, schemas](https://posttrainllm.com/docs/data_inventory/) — [Markdown](https://posttrainllm.com/docs/data_inventory.md) — Reference doc for every dataset wired into posttrainllm — registry entries, what they're for, how to pull them, what the records look like after conversion,…
- [Dataset inventory | PostTrainLLM reference](https://posttrainllm.com/docs/dataset-inventory/) — [Markdown](https://posttrainllm.com/docs/dataset-inventory.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM Decision Log technical documentation](https://posttrainllm.com/docs/decision_log/) — [Markdown](https://posttrainllm.com/docs/decision_log.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Distribution review — 19 September 2026](https://posttrainllm.com/docs/distribution-2026-09-19/) — [Markdown](https://posttrainllm.com/docs/distribution-2026-09-19.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Document Status Registry | PostTrainLLM reference](https://posttrainllm.com/docs/doc-status/) — [Markdown](https://posttrainllm.com/docs/doc-status.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [External Products And Research Reviewed](https://posttrainllm.com/docs/external-products-reviewed/) — [Markdown](https://posttrainllm.com/docs/external-products-reviewed.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Docs | PostTrainLLM reference](https://posttrainllm.com/docs/factory/README/) — [Markdown](https://posttrainllm.com/docs/factory/README.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Artifact Lifecycle Contract | PostTrainLLM reference](https://posttrainllm.com/docs/factory/artifact-lifecycle/) — [Markdown](https://posttrainllm.com/docs/factory/artifact-lifecycle.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Autocorrect adapter recipe and training path](https://posttrainllm.com/docs/factory/autocorrect-adapter-recipe/) — [Markdown](https://posttrainllm.com/docs/factory/autocorrect-adapter-recipe.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Mac-local autocorrect foundation](https://posttrainllm.com/docs/factory/autocorrect-foundation/) — [Markdown](https://posttrainllm.com/docs/factory/autocorrect-foundation.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Autocorrect base-model gate | PostTrainLLM reference](https://posttrainllm.com/docs/factory/autocorrect-model-shortlist/) — [Markdown](https://posttrainllm.com/docs/factory/autocorrect-model-shortlist.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Batch-First Post-Training | PostTrainLLM reference](https://posttrainllm.com/docs/factory/batch-posttraining/) — [Markdown](https://posttrainllm.com/docs/factory/batch-posttraining.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Case Study Template | PostTrainLLM reference](https://posttrainllm.com/docs/factory/case-study-template/) — [Markdown](https://posttrainllm.com/docs/factory/case-study-template.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Enforcement | PostTrainLLM reference](https://posttrainllm.com/docs/factory/enforcement/) — [Markdown](https://posttrainllm.com/docs/factory/enforcement.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Eval Protocol | PostTrainLLM reference](https://posttrainllm.com/docs/factory/eval-protocol/) — [Markdown](https://posttrainllm.com/docs/factory/eval-protocol.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Foundry Evidence Contract — posttrainllm](https://posttrainllm.com/docs/factory/foundry-evidence/) — [Markdown](https://posttrainllm.com/docs/factory/foundry-evidence.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Hugging Face Artifact Storage | PostTrainLLM reference](https://posttrainllm.com/docs/factory/huggingface-artifact-storage/) — [Markdown](https://posttrainllm.com/docs/factory/huggingface-artifact-storage.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [LoRA Geometry Diagnostics | PostTrainLLM reference](https://posttrainllm.com/docs/factory/lora-geometry/) — [Markdown](https://posttrainllm.com/docs/factory/lora-geometry.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Overview | PostTrainLLM reference](https://posttrainllm.com/docs/factory/overview/) — [Markdown](https://posttrainllm.com/docs/factory/overview.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Packaging | PostTrainLLM reference](https://posttrainllm.com/docs/factory/packaging/) — [Markdown](https://posttrainllm.com/docs/factory/packaging.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Post-Training Factory Positioning](https://posttrainllm.com/docs/factory/post-training-factory/) — [Markdown](https://posttrainllm.com/docs/factory/post-training-factory.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Public Artifacts | PostTrainLLM reference](https://posttrainllm.com/docs/factory/public-artifacts/) — [Markdown](https://posttrainllm.com/docs/factory/public-artifacts.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Report Card Cohort and Review | PostTrainLLM reference](https://posttrainllm.com/docs/factory/report-card-cohort/) — [Markdown](https://posttrainllm.com/docs/factory/report-card-cohort.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Fine-Tune Report Card | PostTrainLLM reference](https://posttrainllm.com/docs/factory/report-card/) — [Markdown](https://posttrainllm.com/docs/factory/report-card.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Reports | PostTrainLLM reference](https://posttrainllm.com/docs/factory/reports/) — [Markdown](https://posttrainllm.com/docs/factory/reports.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Run Lifecycle | PostTrainLLM reference](https://posttrainllm.com/docs/factory/run-lifecycle/) — [Markdown](https://posttrainllm.com/docs/factory/run-lifecycle.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory Run Schema | PostTrainLLM reference](https://posttrainllm.com/docs/factory/run-schema/) — [Markdown](https://posttrainllm.com/docs/factory/run-schema.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Decision, forecasting and terminal specialist modes](https://posttrainllm.com/docs/factory/specialist-labs/) — [Markdown](https://posttrainllm.com/docs/factory/specialist-labs.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Feature ideas — interactive learning content](https://posttrainllm.com/docs/feature_ideas/) — [Markdown](https://posttrainllm.com/docs/feature_ideas.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Gallery v1.5 and v2 — build plan](https://posttrainllm.com/docs/gallery_v2_plan/) — [Markdown](https://posttrainllm.com/docs/gallery_v2_plan.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Model guide — building posttrainllm from scratch](https://posttrainllm.com/docs/guides/model_guide/) — [Markdown](https://posttrainllm.com/docs/guides/model_guide.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM study guide technical documentation](https://posttrainllm.com/docs/guides/study_guide/) — [Markdown](https://posttrainllm.com/docs/guides/study_guide.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Training guide | PostTrainLLM reference](https://posttrainllm.com/docs/guides/training_guide/) — [Markdown](https://posttrainllm.com/docs/guides/training_guide.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Qwen3-4B File-Ops Distilled — model card](https://posttrainllm.com/docs/hf-cards/qwen3-4b-file-ops-distilled/) — [Markdown](https://posttrainllm.com/docs/hf-cards/qwen3-4b-file-ops-distilled.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Qwen3-4B ReST Fused — model card](https://posttrainllm.com/docs/hf-cards/qwen3-4b-rest-fused/) — [Markdown](https://posttrainllm.com/docs/hf-cards/qwen3-4b-rest-fused.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Industry learning roadmap | PostTrainLLM reference](https://posttrainllm.com/docs/industry_learning_roadmap/) — [Markdown](https://posttrainllm.com/docs/industry_learning_roadmap.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Using posttrainllm with Continue.dev / Cline / Aider](https://posttrainllm.com/docs/integrations/continue_provider/) — [Markdown](https://posttrainllm.com/docs/integrations/continue_provider.md) — posttrainllm exposes an Ollama-compatible HTTP surface so it drops straight into Continue.dev, Cline, and Aider configs as a local provider. Setup + caveats.
- [Deploying the playground | PostTrainLLM reference](https://posttrainllm.com/docs/integrations/deploy/) — [Markdown](https://posttrainllm.com/docs/integrations/deploy.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [GitHub data integration | PostTrainLLM reference](https://posttrainllm.com/docs/integrations/github_data_integration/) — [Markdown](https://posttrainllm.com/docs/integrations/github_data_integration.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [HuggingFace Datasets Hub integration](https://posttrainllm.com/docs/integrations/hf_datasets_integration/) — [Markdown](https://posttrainllm.com/docs/integrations/hf_datasets_integration.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [lm-evaluation-harness integration for posttrainllm](https://posttrainllm.com/docs/integrations/lm_eval_integration/) — [Markdown](https://posttrainllm.com/docs/integrations/lm_eval_integration.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Model compatibility check (`model-check`)](https://posttrainllm.com/docs/integrations/model-check/) — [Markdown](https://posttrainllm.com/docs/integrations/model-check.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Model run (`model-run`) | PostTrainLLM reference](https://posttrainllm.com/docs/integrations/model-run/) — [Markdown](https://posttrainllm.com/docs/integrations/model-run.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Leaderboard — benchmarks for tiny models](https://posttrainllm.com/docs/leaderboard/) — [Markdown](https://posttrainllm.com/docs/leaderboard.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Learn — course workspace | PostTrainLLM reference](https://posttrainllm.com/docs/learn/) — [Markdown](https://posttrainllm.com/docs/learn.md) — Continue the interactive 13-week inference-systems sprint, with foundations and retained learning references nearby.
- [Learn | PostTrainLLM reference](https://posttrainllm.com/docs/learn/README/) — [Markdown](https://posttrainllm.com/docs/learn/README.md) — posttrainllm's learning corpus — a ground-up curriculum from "what's a neural net" to modern training mechanics, plus reference + research notes.
- [Advanced LLM inference and serving — interview-grade map](https://posttrainllm.com/docs/learn/advanced-llm-inference/) — [Markdown](https://posttrainllm.com/docs/learn/advanced-llm-inference.md) — Senior/staff inference-optimization interview topics — KV cache & paging, batching, speculative decoding, quantization, attention kernels/variants, long…
- [Advanced LLM training and post-training — interview-grade…](https://posttrainllm.com/docs/learn/advanced-llm-training/) — [Markdown](https://posttrainllm.com/docs/learn/advanced-llm-training.md) — Senior/staff interview topics for LLM training at scale (parallelism, precision, MoE, data) and post-training (RLHF/DPO/GRPO/reward modeling/distillation),…
- [Advanced architecture, RAG/agents, eval and ML system…](https://posttrainllm.com/docs/learn/advanced-ml-systems-eval/) — [Markdown](https://posttrainllm.com/docs/learn/advanced-ml-systems-eval.md) — Senior/staff topics spanning modern transformer architecture, the linear-algebra↔transformers whiteboard, RAG, agents, evaluation (LLM-as-judge, perplexity,…
- [Agent context as a memory hierarchy](https://posttrainllm.com/docs/learn/agent-context-hierarchy/) — [Markdown](https://posttrainllm.com/docs/learn/agent-context-hierarchy.md) — The L1/L2/L3 framing for agent context engineering — what we stole from the Shortcut vertical-agents essay and where each steal lives in this repo.
- [ANE research dossier for posttrainllm](https://posttrainllm.com/docs/learn/ane-research/dossier/) — [Markdown](https://posttrainllm.com/docs/learn/ane-research/dossier.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [M6 — ANE Bisect Findings (2026-06-08)](https://posttrainllm.com/docs/learn/ane-research/m6-findings/) — [Markdown](https://posttrainllm.com/docs/learn/ane-research/m6-findings.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [App Intents comparison for Planner v7](https://posttrainllm.com/docs/learn/app-intents-comparison/) — [Markdown](https://posttrainllm.com/docs/learn/app-intents-comparison.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Apple on-device Foundation Models — what they are, and…](https://posttrainllm.com/docs/learn/apple-on-device-foundation-models/) — [Markdown](https://posttrainllm.com/docs/learn/apple-on-device-foundation-models.md) — What Apple's FoundationModels framework + on-device model give you, the bridge we built to score it on our gates, the measured verdict (can't ground actions;…
- [Session 1 — What's a neural net (the building block)](https://posttrainllm.com/docs/learn/archive/session-01-neural-net-basics-dense/) — [Markdown](https://posttrainllm.com/docs/learn/archive/session-01-neural-net-basics-dense.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [The buildable AI journey | PostTrainLLM reference](https://posttrainllm.com/docs/learn/artifact-journey/) — [Markdown](https://posttrainllm.com/docs/learn/artifact-journey.md) — A nine-stage journey from transformer foundations to local specialist models, agents, packages, runtime decisions, and causal probes.
- [Castform's RL fine-tune platform — what we stole](https://posttrainllm.com/docs/learn/castform-rl-finetune/) — [Markdown](https://posttrainllm.com/docs/learn/castform-rl-finetune.md) — Patterns lifted from castform.com (RL fine-tune SaaS) into posttrainllm's specialist-training surface. Sibling page to docs/learn/agent-context-hierarchy.md.
- [Competitive landscape — fine-tune, eval, interpretability](https://posttrainllm.com/docs/learn/competitive-landscape/) — [Markdown](https://posttrainllm.com/docs/learn/competitive-landscape.md) — A 2026 map of the products around "make a model good at your task" and "is my agent any good," with the Mac-first whitespace called out.…
- [Learning coverage map — every subsystem has a home](https://posttrainllm.com/docs/learn/coverage-map/) — [Markdown](https://posttrainllm.com/docs/learn/coverage-map.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM learning curriculum — ground up](https://posttrainllm.com/docs/learn/curriculum/) — [Markdown](https://posttrainllm.com/docs/learn/curriculum.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Diversity-driven small-model reasoning (Spectrum-to-Signal)](https://posttrainllm.com/docs/learn/diversity-driven-small-model-reasoning/) — [Markdown](https://posttrainllm.com/docs/learn/diversity-driven-small-model-reasoning.md) — Three transferable training methods from the VibeThinker papers — Diversity-Exploring Distillation, MaxEnt-Guided Policy Optimization (MGPO), and specialist…
- [Encoder-decoder adapters: LoRA on a seq2seq model for text…](https://posttrainllm.com/docs/learn/encoder-decoder-adapters/) — [Markdown](https://posttrainllm.com/docs/learn/encoder-decoder-adapters.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Mathematically essential vs engineering optimization](https://posttrainllm.com/docs/learn/essential-vs-optimization/) — [Markdown](https://posttrainllm.com/docs/learn/essential-vs-optimization.md) — The single most useful lens on this project — which parts compute the model's function (irreducible math, the correctness oracle) vs which parts only make…
- [Eval matrix — what every Pace LoRA actually scores](https://posttrainllm.com/docs/learn/eval-matrix-2026-06-08/) — [Markdown](https://posttrainllm.com/docs/learn/eval-matrix-2026-06-08.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Eval methodology — the gate finding](https://posttrainllm.com/docs/learn/eval-methodology-2026-06-08/) — [Markdown](https://posttrainllm.com/docs/learn/eval-methodology-2026-06-08.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [External references — articles worth reading](https://posttrainllm.com/docs/learn/external-references/) — [Markdown](https://posttrainllm.com/docs/learn/external-references.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Halo: source review and Mac-local disposition](https://posttrainllm.com/docs/learn/halo/) — [Markdown](https://posttrainllm.com/docs/learn/halo.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Inference systems — 13-week owner sprint](https://posttrainllm.com/docs/learn/inference-systems-13w/) — [Markdown](https://posttrainllm.com/docs/learn/inference-systems-13w.md) — 28 September–27 December 2026 route through transformer mechanics, measurement, kernels, serving, and the single-device-to-cluster boundary.
- [Learning journal | PostTrainLLM reference](https://posttrainllm.com/docs/learn/journal/) — [Markdown](https://posttrainllm.com/docs/learn/journal.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Legacy walkthrough — Python, WASM, and WebGPU](https://posttrainllm.com/docs/learn/legacy-walkthrough/) — [Markdown](https://posttrainllm.com/docs/learn/legacy-walkthrough.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [How to Cook an LLM: lesson and integration notes](https://posttrainllm.com/docs/learn/llm-kitchen/) — [Markdown](https://posttrainllm.com/docs/learn/llm-kitchen.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [How an LLM actually works (the matmul-first explanation)](https://posttrainllm.com/docs/learn/llm-mechanics-fundamentals/) — [Markdown](https://posttrainllm.com/docs/learn/llm-mechanics-fundamentals.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Mac-local AI mastery map — what's buildable, what's built,…](https://posttrainllm.com/docs/learn/mac-mastery-map/) — [Markdown](https://posttrainllm.com/docs/learn/mac-mastery-map.md) — The learning + build agenda for mastering Mac-local AI. Every capability buildable on a single Mac, annotated with fleet coverage (built / partial /…
- [Model vs agent — what's actually different](https://posttrainllm.com/docs/learn/model-vs-agent/) — [Markdown](https://posttrainllm.com/docs/learn/model-vs-agent.md) — The architectural distinction between a fine-tuned model and the agent that runs it, and what "making a good model into an agent" actually requires.…
- [PostTrainLLM — study queue technical documentation](https://posttrainllm.com/docs/learn/new-things/) — [Markdown](https://posttrainllm.com/docs/learn/new-things.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Open Kernel Challenge | PostTrainLLM reference](https://posttrainllm.com/docs/learn/openkernel/) — [Markdown](https://posttrainllm.com/docs/learn/openkernel.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Qwen3-VL mRoPE + DeepStack — math spec from HF reference](https://posttrainllm.com/docs/learn/qwen3-vl-mrope-deepstack/) — [Markdown](https://posttrainllm.com/docs/learn/qwen3-vl-mrope-deepstack.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Reproducing Qwen Chess under a 50M-parameter ceiling](https://posttrainllm.com/docs/learn/reproducing-qwen-chess-under-50m/) — [Markdown](https://posttrainllm.com/docs/learn/reproducing-qwen-chess-under-50m.md) — What the external Qwen Chess project actually proves, what posttrainllm can reproduce, and where the 8B-to-44M compression experiment begins.
- [RL + multi-autonomous-agents roadmap — the AI game as a…](https://posttrainllm.com/docs/learn/rl-multi-agent-roadmap/) — [Markdown](https://posttrainllm.com/docs/learn/rl-multi-agent-roadmap.md) — Parked blueprint. The fleet's AI game (autonomous characters in a world) is already a multi-agent RL environment — the scarcest RL ingredient. This is the…
- [Session 1 — From a line to a learned line](https://posttrainllm.com/docs/learn/session-01-neural-net-basics/) — [Markdown](https://posttrainllm.com/docs/learn/session-01-neural-net-basics.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 2 — How the search for `m` and `b` actually works](https://posttrainllm.com/docs/learn/session-02-gradient-descent/) — [Markdown](https://posttrainllm.com/docs/learn/session-02-gradient-descent.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 3 — What makes a neural net more than linear…](https://posttrainllm.com/docs/learn/session-03-non-linearities/) — [Markdown](https://posttrainllm.com/docs/learn/session-03-non-linearities.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 4 — The taxonomy of ML approaches (and where…](https://posttrainllm.com/docs/learn/session-04-ml-paradigms/) — [Markdown](https://posttrainllm.com/docs/learn/session-04-ml-paradigms.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 5 — Scaling: why bigger models know more (and what…](https://posttrainllm.com/docs/learn/session-05-scaling/) — [Markdown](https://posttrainllm.com/docs/learn/session-05-scaling.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 6 — Tokenization + embeddings: how text becomes…](https://posttrainllm.com/docs/learn/session-06-tokenization-embeddings/) — [Markdown](https://posttrainllm.com/docs/learn/session-06-tokenization-embeddings.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 7 — How models learn behavior (supervised,…](https://posttrainllm.com/docs/learn/session-07-behavior-learning/) — [Markdown](https://posttrainllm.com/docs/learn/session-07-behavior-learning.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 8 — Training mechanics: the actual loop](https://posttrainllm.com/docs/learn/session-08-training-mechanics/) — [Markdown](https://posttrainllm.com/docs/learn/session-08-training-mechanics.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 9 — Vectors, matrices, tensors: the shapes that…](https://posttrainllm.com/docs/learn/session-09-tensors/) — [Markdown](https://posttrainllm.com/docs/learn/session-09-tensors.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 10 — Attention and transformer blocks: routing…](https://posttrainllm.com/docs/learn/session-10-attention/) — [Markdown](https://posttrainllm.com/docs/learn/session-10-attention.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session 11 — Evals, rewards, and self-improvement: making…](https://posttrainllm.com/docs/learn/session-11-evals-rewards/) — [Markdown](https://posttrainllm.com/docs/learn/session-11-evals-rewards.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Small-model tool-calling: the SOTA playbook (what others do)](https://posttrainllm.com/docs/learn/small-model-tool-calling-playbook/) — [Markdown](https://posttrainllm.com/docs/learn/small-model-tool-calling-playbook.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Speech and systems topics — interview-grade map](https://posttrainllm.com/docs/learn/speech-and-systems-topics/) — [Markdown](https://posttrainllm.com/docs/learn/speech-and-systems-topics.md) — Voice-pipeline latency, WER, speech-to-speech, fine-tuning debugging, feature selection, queues vs websockets, FSDP2 — each mapped to the best external source…
- [Tool-calling: how close can a Mac-local small model get to…](https://posttrainllm.com/docs/learn/tool-calling-frontier-parity/) — [Markdown](https://posttrainllm.com/docs/learn/tool-calling-frontier-parity.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [The WebGPU execution model (for reading our shaders)](https://posttrainllm.com/docs/learn/webgpu-execution-model/) — [Markdown](https://posttrainllm.com/docs/learn/webgpu-execution-model.md) — The mental model you need to read posttrainllm's WGSL compute kernels and explain them to another engineer — device/queue, pipeline, dispatch, workgroups,…
- [Learning Pipeline | PostTrainLLM reference](https://posttrainllm.com/docs/learning-pipeline/) — [Markdown](https://posttrainllm.com/docs/learning-pipeline.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Learning Progress Tracker | PostTrainLLM reference](https://posttrainllm.com/docs/learning-progress/) — [Markdown](https://posttrainllm.com/docs/learning-progress.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Document library — where to look](https://posttrainllm.com/docs/library/) — [Markdown](https://posttrainllm.com/docs/library.md) — A labeled map of the retained learning, reference, factory, evidence, and historical documentation.
- [Native macOS app — build plan | PostTrainLLM reference](https://posttrainllm.com/docs/mac_app_plan/) — [Markdown](https://posttrainllm.com/docs/mac_app_plan.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM — a detailed write-up](https://posttrainllm.com/docs/notes/) — [Markdown](https://posttrainllm.com/docs/notes.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pace landing page — pattern mine + draft](https://posttrainllm.com/docs/pace-landing-draft/) — [Markdown](https://posttrainllm.com/docs/pace-landing-draft.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pace model manifest — every model role, decided and on disk](https://posttrainllm.com/docs/pace-model-manifest/) — [Markdown](https://posttrainllm.com/docs/pace-model-manifest.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Parked Lanes | PostTrainLLM reference](https://posttrainllm.com/docs/parked/README/) — [Markdown](https://posttrainllm.com/docs/parked/README.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Parked: ANE and CoreML | PostTrainLLM reference](https://posttrainllm.com/docs/parked/ane-coreml/) — [Markdown](https://posttrainllm.com/docs/parked/ane-coreml.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Parked: Broad App Polish | PostTrainLLM reference](https://posttrainllm.com/docs/parked/app-polish/) — [Markdown](https://posttrainllm.com/docs/parked/app-polish.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Parked: Browser and WebGPU | PostTrainLLM reference](https://posttrainllm.com/docs/parked/browser/) — [Markdown](https://posttrainllm.com/docs/parked/browser.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Parked: Tier 5 Research | PostTrainLLM reference](https://posttrainllm.com/docs/parked/tier5/) — [Markdown](https://posttrainllm.com/docs/parked/tier5.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Parked: VLM | PostTrainLLM reference](https://posttrainllm.com/docs/parked/vlm/) — [Markdown](https://posttrainllm.com/docs/parked/vlm.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [First Worked Run — `posttrainllm bench`](https://posttrainllm.com/docs/performance/benchmark_first_run/) — [Markdown](https://posttrainllm.com/docs/performance/benchmark_first_run.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM Inference Benchmark Harness — Design](https://posttrainllm.com/docs/performance/benchmark_harness_design/) — [Markdown](https://posttrainllm.com/docs/performance/benchmark_harness_design.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Cold-start bundle — results | PostTrainLLM reference](https://posttrainllm.com/docs/performance/cold_start_results/) — [Markdown](https://posttrainllm.com/docs/performance/cold_start_results.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [CPU speedup bundle — measured results](https://posttrainllm.com/docs/performance/cpu_speedup_results/) — [Markdown](https://posttrainllm.com/docs/performance/cpu_speedup_results.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [CPU utilization research — what we're leaving on the table](https://posttrainllm.com/docs/performance/cpu_utilization_research/) — [Markdown](https://posttrainllm.com/docs/performance/cpu_utilization_research.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Data-side performance levers | PostTrainLLM reference](https://posttrainllm.com/docs/performance/data_perf/) — [Markdown](https://posttrainllm.com/docs/performance/data_perf.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Determinism contract | PostTrainLLM reference](https://posttrainllm.com/docs/performance/determinism/) — [Markdown](https://posttrainllm.com/docs/performance/determinism.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Flash Attention 2 — backward pass notes](https://posttrainllm.com/docs/performance/fa2_backward_notes/) — [Markdown](https://posttrainllm.com/docs/performance/fa2_backward_notes.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Flash Attention 2 — forward pass notes](https://posttrainllm.com/docs/performance/fa2_forward_notes/) — [Markdown](https://posttrainllm.com/docs/performance/fa2_forward_notes.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Gradient checkpointing — worked example](https://posttrainllm.com/docs/performance/gradient_checkpointing_results/) — [Markdown](https://posttrainllm.com/docs/performance/gradient_checkpointing_results.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [KV Cache Optimization: GQA Audit, In-Place Buffers,…](https://posttrainllm.com/docs/performance/kv_cache_optimization/) — [Markdown](https://posttrainllm.com/docs/performance/kv_cache_optimization.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Memory tradeoffs — bf16, gradient accumulation, gradient…](https://posttrainllm.com/docs/performance/memory_tradeoffs/) — [Markdown](https://posttrainllm.com/docs/performance/memory_tradeoffs.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Online softmax in attention — why and how](https://posttrainllm.com/docs/performance/online_softmax_in_attention/) — [Markdown](https://posttrainllm.com/docs/performance/online_softmax_in_attention.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [MLXFast SDPA and tied-embedding audit (2026-05-30)](https://posttrainllm.com/docs/performance/perf_audit_mlxfast_tied/) — [Markdown](https://posttrainllm.com/docs/performance/perf_audit_mlxfast_tied.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [The browser-frontier performance quest](https://posttrainllm.com/docs/performance/perf_quest/) — [Markdown](https://posttrainllm.com/docs/performance/perf_quest.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Performance research — what's done, what's plausibly next,…](https://posttrainllm.com/docs/performance/perf_research/) — [Markdown](https://posttrainllm.com/docs/performance/perf_research.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Performance notes | PostTrainLLM reference](https://posttrainllm.com/docs/performance/performance/) — [Markdown](https://posttrainllm.com/docs/performance/performance.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [YOCO — "You Only Cache Once" worked example](https://posttrainllm.com/docs/performance/yoco_results/) — [Markdown](https://posttrainllm.com/docs/performance/yoco_results.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [5.1 reasoning training on a 22M model](https://posttrainllm.com/docs/prds/1-reasoning-on-22M/) — [Markdown](https://posttrainllm.com/docs/prds/1-reasoning-on-22M.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [5.2 test-time compute scaling experiment](https://posttrainllm.com/docs/prds/2-testtime-compute-scaling/) — [Markdown](https://posttrainllm.com/docs/prds/2-testtime-compute-scaling.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [5.3 vision-language toy | PostTrainLLM reference](https://posttrainllm.com/docs/prds/3-vision-language-toy/) — [Markdown](https://posttrainllm.com/docs/prds/3-vision-language-toy.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [5.4 diffusion LM micro-implementation](https://posttrainllm.com/docs/prds/4-diffusion-lm-micro/) — [Markdown](https://posttrainllm.com/docs/prds/4-diffusion-lm-micro.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [5.5 real sparse MoE kernels (Metal)](https://posttrainllm.com/docs/prds/5-sparse-moe-kernels/) — [Markdown](https://posttrainllm.com/docs/prds/5-sparse-moe-kernels.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [5.6 TTS toy (audio-token GPT) | PostTrainLLM reference](https://posttrainllm.com/docs/prds/6-tts-toy/) — [Markdown](https://posttrainllm.com/docs/prds/6-tts-toy.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [5.7 specialized explainer-video model (Lamina-like)](https://posttrainllm.com/docs/prds/7-explainer-video-model/) — [Markdown](https://posttrainllm.com/docs/prds/7-explainer-video-model.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [A1 first specialist end-to-end (tool-caller)](https://posttrainllm.com/docs/prds/A1-first-specialist-tool-caller/) — [Markdown](https://posttrainllm.com/docs/prds/A1-first-specialist-tool-caller.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B1 second specialist (SQL POC)](https://posttrainllm.com/docs/prds/B1-second-specialist-shell-or-sql/) — [Markdown](https://posttrainllm.com/docs/prds/B1-second-specialist-shell-or-sql.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B10 quality classifier on pretrain data](https://posttrainllm.com/docs/prds/B10-quality-classifier/) — [Markdown](https://posttrainllm.com/docs/prds/B10-quality-classifier.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B11 WSD (warmup-stable-decay) schedule](https://posttrainllm.com/docs/prds/B11-wsd-schedule/) — [Markdown](https://posttrainllm.com/docs/prds/B11-wsd-schedule.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B12 loss-spike recovery + replay](https://posttrainllm.com/docs/prds/B12-loss-spike-recovery/) — [Markdown](https://posttrainllm.com/docs/prds/B12-loss-spike-recovery.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B13 interp-on-checkpoints methodology](https://posttrainllm.com/docs/prds/B13-interp-on-checkpoints/) — [Markdown](https://posttrainllm.com/docs/prds/B13-interp-on-checkpoints.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B14 speculative decoding (Mini-Llama draft for Mega target)](https://posttrainllm.com/docs/prds/B14-speculative-decoding/) — [Markdown](https://posttrainllm.com/docs/prds/B14-speculative-decoding.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B15 layer-wise LR decay for SFT](https://posttrainllm.com/docs/prds/B15-layerwise-lr-decay-sft/) — [Markdown](https://posttrainllm.com/docs/prds/B15-layerwise-lr-decay-sft.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B16 M5 Neural Accelerator prefill benchmark](https://posttrainllm.com/docs/prds/B16-m5-na-prefill-bench/) — [Markdown](https://posttrainllm.com/docs/prds/B16-m5-na-prefill-bench.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B17 SAE Lens interop / Neuronpedia format export](https://posttrainllm.com/docs/prds/B17-saelens-interop/) — [Markdown](https://posttrainllm.com/docs/prds/B17-saelens-interop.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B18 nanochat-style `--depth` single-knob HP derivation](https://posttrainllm.com/docs/prds/B18-nanochat-depth-knob/) — [Markdown](https://posttrainllm.com/docs/prds/B18-nanochat-depth-knob.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B19 Group-SAE (layer-group SAE training)](https://posttrainllm.com/docs/prds/B19-group-sae/) — [Markdown](https://posttrainllm.com/docs/prds/B19-group-sae.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B2/B2b/B3/B7 mini-router family (bundled)](https://posttrainllm.com/docs/prds/B2-B7-router-family/) — [Markdown](https://posttrainllm.com/docs/prds/B2-B7-router-family.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B21 Micro-AutoMixer for specialist data mixes](https://posttrainllm.com/docs/prds/B21-micro-automixer/) — [Markdown](https://posttrainllm.com/docs/prds/B21-micro-automixer.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B23 agent eval protocol hardening](https://posttrainllm.com/docs/prds/B23-agent-eval-protocol/) — [Markdown](https://posttrainllm.com/docs/prds/B23-agent-eval-protocol.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B25 ScaleDown Challenge specialist](https://posttrainllm.com/docs/prds/B25-scaledown-specialist/) — [Markdown](https://posttrainllm.com/docs/prds/B25-scaledown-specialist.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B26 server-side deferred tools](https://posttrainllm.com/docs/prds/B26-deferred-tools/) — [Markdown](https://posttrainllm.com/docs/prds/B26-deferred-tools.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B28 composite reward framework](https://posttrainllm.com/docs/prds/B28-composite-reward-framework/) — [Markdown](https://posttrainllm.com/docs/prds/B28-composite-reward-framework.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B31 unified model gallery + project-level model pins](https://posttrainllm.com/docs/prds/B31-gallery-and-project-pins/) — [Markdown](https://posttrainllm.com/docs/prds/B31-gallery-and-project-pins.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B32 `posttrainllm eval` as a CI / pre-commit gate](https://posttrainllm.com/docs/prds/B32-eval-ci-gate/) — [Markdown](https://posttrainllm.com/docs/prds/B32-eval-ci-gate.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B33 one-command laptop-finetune onboarding](https://posttrainllm.com/docs/prds/B33-laptop-finetune-onboarding/) — [Markdown](https://posttrainllm.com/docs/prds/B33-laptop-finetune-onboarding.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B34 batched eval-runtime + pluggable MLX backend (oMLX…](https://posttrainllm.com/docs/prds/B34-batched-eval-runtime/) — [Markdown](https://posttrainllm.com/docs/prds/B34-batched-eval-runtime.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B35 local-agent vertical PoC — code reviewer on a Mac](https://posttrainllm.com/docs/prds/B35-local-agent-vertical-poc/) — [Markdown](https://posttrainllm.com/docs/prds/B35-local-agent-vertical-poc.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B5 cloud-escalation training signal](https://posttrainllm.com/docs/prds/B5-cloud-escalate-training/) — [Markdown](https://posttrainllm.com/docs/prds/B5-cloud-escalate-training.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B6 Mac app demo | PostTrainLLM reference](https://posttrainllm.com/docs/prds/B6-mac-app-demo/) — [Markdown](https://posttrainllm.com/docs/prds/B6-mac-app-demo.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B8 multilingual specialist (Indic-focus first)](https://posttrainllm.com/docs/prds/B8-multilingual-specialist/) — [Markdown](https://posttrainllm.com/docs/prds/B8-multilingual-specialist.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B9 energy J/token measurement | PostTrainLLM reference](https://posttrainllm.com/docs/prds/B9-energy-per-token/) — [Markdown](https://posttrainllm.com/docs/prds/B9-energy-per-token.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [C10 training-run dashboard | PostTrainLLM reference](https://posttrainllm.com/docs/prds/C10-train-run-dashboard/) — [Markdown](https://posttrainllm.com/docs/prds/C10-train-run-dashboard.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [C3 DoRA on-disk adapter format](https://posttrainllm.com/docs/prds/C3-dora-ondisk-format/) — [Markdown](https://posttrainllm.com/docs/prds/C3-dora-ondisk-format.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [C4 tool-call extractor — BPE tokenizer support](https://posttrainllm.com/docs/prds/C4-tool-extractor-bpe/) — [Markdown](https://posttrainllm.com/docs/prds/C4-tool-extractor-bpe.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [C5 decode jitter under thermal load](https://posttrainllm.com/docs/prds/C5-decode-jitter-thermal/) — [Markdown](https://posttrainllm.com/docs/prds/C5-decode-jitter-thermal.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [C9 determinism harness — bit-exact replay of step N](https://posttrainllm.com/docs/prds/C9-determinism-harness/) — [Markdown](https://posttrainllm.com/docs/prds/C9-determinism-harness.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [E6 `posttrainllm eval-scaledown` harness](https://posttrainllm.com/docs/prds/E6-eval-scaledown/) — [Markdown](https://posttrainllm.com/docs/prds/E6-eval-scaledown.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [GPU / research backlog — what's left, and how to do it](https://posttrainllm.com/docs/prds/GPU-RESEARCH-BACKLOG/) — [Markdown](https://posttrainllm.com/docs/prds/GPU-RESEARCH-BACKLOG.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PRD Priority Triage | PostTrainLLM reference](https://posttrainllm.com/docs/prds/PRIORITY/) — [Markdown](https://posttrainllm.com/docs/prds/PRIORITY.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [docs/prds/ — Product Requirement Briefs](https://posttrainllm.com/docs/prds/README/) — [Markdown](https://posttrainllm.com/docs/prds/README.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PRD status — code-verified audit (2026-06-20)](https://posttrainllm.com/docs/prds/STATUS/) — [Markdown](https://posttrainllm.com/docs/prds/STATUS.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Capability retention under fine-tuning (PRD — backlog)](https://posttrainllm.com/docs/prds/capability-retention/) — [Markdown](https://posttrainllm.com/docs/prds/capability-retention.md) — Measure and preserve a small model's GENERAL intelligence when we specialize it (tool-calling, style, persona). This session proved specialization erodes…
- [PRD — On-device continual-learning loop](https://posttrainllm.com/docs/prds/continual-learning-loop/) — [Markdown](https://posttrainllm.com/docs/prds/continual-learning-loop.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Everyday specialist benchmark | PostTrainLLM reference](https://posttrainllm.com/docs/prds/everyday-specialist-benchmark/) — [Markdown](https://posttrainllm.com/docs/prds/everyday-specialist-benchmark.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Factory completeness tracker — meta-PRD](https://posttrainllm.com/docs/prds/factory-completeness-tracker/) — [Markdown](https://posttrainllm.com/docs/prds/factory-completeness-tracker.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Planner v7 — tools-in-prompt (generalizable function…](https://posttrainllm.com/docs/prds/factory-planner-v7-tools-in-prompt/) — [Markdown](https://posttrainllm.com/docs/prds/factory-planner-v7-tools-in-prompt.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [VLM M4 architecture decision — Qwen3-VL port vs LLaVA…](https://posttrainllm.com/docs/prds/factory-vision-m4-architecture-decision/) — [Markdown](https://posttrainllm.com/docs/prds/factory-vision-m4-architecture-decision.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [VLM M4 — Qwen3-VL implementation plan](https://posttrainllm.com/docs/prds/factory-vision-m4-impl-plan/) — [Markdown](https://posttrainllm.com/docs/prds/factory-vision-m4-impl-plan.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Vision-language specialist for Pace screen reading](https://posttrainllm.com/docs/prds/factory-vision-specialist/) — [Markdown](https://posttrainllm.com/docs/prds/factory-vision-specialist.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Game-as-RL-environment PoC (PRD)](https://posttrainllm.com/docs/prds/game-rl-environment-poc/) — [Markdown](https://posttrainllm.com/docs/prds/game-rl-environment-poc.md) — Turn the fleet's AI game (autonomous characters in a world) into a Mac-local RL environment and train a self-improving NPC with GRPO. Environments are the…
- [GEPA prompt evolution — automated system-prompt iteration](https://posttrainllm.com/docs/prds/gepa-prompt-evolution/) — [Markdown](https://posttrainllm.com/docs/prds/gepa-prompt-evolution.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Local-model arena — turn-based strategy, self-play RLVR…](https://posttrainllm.com/docs/prds/local-model-arena-selfplay/) — [Markdown](https://posttrainllm.com/docs/prds/local-model-arena-selfplay.md) — Pit Mac-local models against each other and against frontier models in turn-based strategy games, then self-play-RL a local model until it beats a frontier…
- [Port macOS 26 int8 direct ANE array handoff into M8 chain](https://posttrainllm.com/docs/prds/macos26-int8-ane-handoff-port/) — [Markdown](https://posttrainllm.com/docs/prds/macos26-int8-ane-handoff-port.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Multi-turn / agentic tool-calling eval (PRD)](https://posttrainllm.com/docs/prds/multi-turn-agentic-eval/) — [Markdown](https://posttrainllm.com/docs/prds/multi-turn-agentic-eval.md) — A Mac-local, frontier-gated, STATEFUL multi-turn eval — the single biggest blind spot in our tool-calling work. Single-turn numbers (88.7 for our best 4B)…
- [Pace planner v11 — ship gate | PostTrainLLM reference](https://posttrainllm.com/docs/prds/pace-planner-v11-ship-gate/) — [Markdown](https://posttrainllm.com/docs/prds/pace-planner-v11-ship-gate.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pace planner v11 — unhappy-path training data PRD](https://posttrainllm.com/docs/prds/pace-planner-v11-training-data/) — [Markdown](https://posttrainllm.com/docs/prds/pace-planner-v11-training-data.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pace Task Loop v1 — long-horizon tasks + the bulk-download…](https://posttrainllm.com/docs/prds/pace-task-loop-v1/) — [Markdown](https://posttrainllm.com/docs/prds/pace-task-loop-v1.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [QLoRA on Mac — fine-tuning larger open models on 48 GB](https://posttrainllm.com/docs/prds/qlora-large-model-finetune/) — [Markdown](https://posttrainllm.com/docs/prds/qlora-large-model-finetune.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Quantized inference in posttrainllm's HFModel — Swift-side…](https://posttrainllm.com/docs/prds/quantized-inference-swift/) — [Markdown](https://posttrainllm.com/docs/prds/quantized-inference-swift.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Self-improving agents — the closed loop (PRD)](https://posttrainllm.com/docs/prds/self-improving-agents/) — [Markdown](https://posttrainllm.com/docs/prds/self-improving-agents.md) — A verifiable act, score, learn, and curriculum loop for testing whether a Mac-local agent can improve without a teacher.
- [Specialist capability graph and verified cascade](https://posttrainllm.com/docs/prds/specialist-capability-graph/) — [Markdown](https://posttrainllm.com/docs/prds/specialist-capability-graph.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pace planner specialist — first factory customer](https://posttrainllm.com/docs/prds/specialist-pace-planner/) — [Markdown](https://posttrainllm.com/docs/prds/specialist-pace-planner.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM product thesis — the embedded local-AI…](https://posttrainllm.com/docs/prds/tinygpt-product-thesis/) — [Markdown](https://posttrainllm.com/docs/prds/tinygpt-product-thesis.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [VLM A/B — UI-Venus-1.5-2B vs Qwen3-VL-2B for Pace's vision…](https://posttrainllm.com/docs/prds/vlm-ab-uivenus-vs-qwen3vl/) — [Markdown](https://posttrainllm.com/docs/prds/vlm-ab-uivenus-vs-qwen3vl.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Progress dashboard — moved | PostTrainLLM reference](https://posttrainllm.com/docs/progress/) — [Markdown](https://posttrainllm.com/docs/progress.md) — Consolidated into docs/PLAN.md §1 (shipped). Headline metrics live there.
- [Quality Bundle: tests, lint, crash recovery](https://posttrainllm.com/docs/quality_bundle/) — [Markdown](https://posttrainllm.com/docs/quality_bundle.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [`posttrainllm quickstart` — your first specialist in one…](https://posttrainllm.com/docs/quickstart/) — [Markdown](https://posttrainllm.com/docs/quickstart.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM Recipes technical documentation](https://posttrainllm.com/docs/recipes/README/) — [Markdown](https://posttrainllm.com/docs/recipes/README.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Recipe — micro-AutoMixer for data-mix ratios (B21)](https://posttrainllm.com/docs/recipes/automix/) — [Markdown](https://posttrainllm.com/docs/recipes/automix.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Recipe — B25 ScaleDown specialist (extractive context…](https://posttrainllm.com/docs/recipes/b25-scaledown/) — [Markdown](https://posttrainllm.com/docs/recipes/b25-scaledown.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Balanced training mix via reasoning-depth classification](https://posttrainllm.com/docs/recipes/balanced-training-mix/) — [Markdown](https://posttrainllm.com/docs/recipes/balanced-training-mix.md) — Use B30 ahead of B29 to control single-hop / multi-hop / comparison ratios in the SFT corpus.
- [Cookbook - Character Specialist](https://posttrainllm.com/docs/recipes/cookbook-character-specialist/) — [Markdown](https://posttrainllm.com/docs/recipes/cookbook-character-specialist.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Cookbook - Personal Code Specialist](https://posttrainllm.com/docs/recipes/cookbook-personal-code-specialist/) — [Markdown](https://posttrainllm.com/docs/recipes/cookbook-personal-code-specialist.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Cookbook - posttrainllm with Pydantic AI](https://posttrainllm.com/docs/recipes/cookbook-pydantic-ai/) — [Markdown](https://posttrainllm.com/docs/recipes/cookbook-pydantic-ai.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Cookbook - posttrainllm with smolagents](https://posttrainllm.com/docs/recipes/cookbook-smolagents/) — [Markdown](https://posttrainllm.com/docs/recipes/cookbook-smolagents.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Recipe: distill a big model into a small local specialist](https://posttrainllm.com/docs/recipes/distill-specialist/) — [Markdown](https://posttrainllm.com/docs/recipes/distill-specialist.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Recipe — distill function-calling from Phi-3-mini →…](https://posttrainllm.com/docs/recipes/distillation-fc/) — [Markdown](https://posttrainllm.com/docs/recipes/distillation-fc.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Eval gate (CI / pre-commit) | PostTrainLLM reference](https://posttrainllm.com/docs/recipes/eval-gate/) — [Markdown](https://posttrainllm.com/docs/recipes/eval-gate.md) — Gate a posttrainllm specialist in CI — `posttrainllm eval-gate` exits non-zero when any declared eval suite regresses past threshold, on a self-hosted Mac…
- [Recipe - Pace Planner Eval Protocol](https://posttrainllm.com/docs/recipes/eval_planner/) — [Markdown](https://posttrainllm.com/docs/recipes/eval_planner.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [From .atraj rollouts to a trained specialist](https://posttrainllm.com/docs/recipes/from-traces/) — [Markdown](https://posttrainllm.com/docs/recipes/from-traces.md) — B22 + B29 + B30 + posttrainllm sft chained — the closed substrate→training loop, V1.
- [MLX Export | PostTrainLLM reference](https://posttrainllm.com/docs/recipes/mlx-export/) — [Markdown](https://posttrainllm.com/docs/recipes/mlx-export.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Recipe — Pace planner specialist](https://posttrainllm.com/docs/recipes/pace-planner/) — [Markdown](https://posttrainllm.com/docs/recipes/pace-planner.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B20 — Learnable cross-stream attention evaluation](https://posttrainllm.com/docs/research/cross-stream-attention-evaluation/) — [Markdown](https://posttrainllm.com/docs/research/cross-stream-attention-evaluation.md) — Read-and-evaluate write-up on the modded-nanogpt speedrun's "cross-stream attention" trick; decide whether to adopt for posttrainllm.
- [Indic-language evals — MILU + IndicGenBench wiring](https://posttrainllm.com/docs/research/indic_evals/) — [Markdown](https://posttrainllm.com/docs/research/indic_evals.md) — How posttrainllm scores Indic-language ability — MILU multi-choice across 11 langs, IndicGenBench XQuAD extractive QA. The Wave 4 gate before claiming Hindi…
- [State of LLM Inference Benchmarks (May 2026)](https://posttrainllm.com/docs/research/inference_benchmarks_may_2026/) — [Markdown](https://posttrainllm.com/docs/research/inference_benchmarks_may_2026.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Mac inference baseline — M5 Pro / 48GB](https://posttrainllm.com/docs/research/mac_decode_baseline_m5pro/) — [Markdown](https://posttrainllm.com/docs/research/mac_decode_baseline_m5pro.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Mac SLM agentic leaderboard v0](https://posttrainllm.com/docs/research/mac_slm_leaderboard_v0/) — [Markdown](https://posttrainllm.com/docs/research/mac_slm_leaderboard_v0.md) — One artifact that cross-cuts decode speed, BFCL, τ-bench, and Pace unhappy-paths — the publication-shape view we're missing.
- [LLM Quality Benchmark Landscape, May 2026 — Survey for…](https://posttrainllm.com/docs/research/quality_benchmarks_may_2026/) — [Markdown](https://posttrainllm.com/docs/research/quality_benchmarks_may_2026.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Wave 2.5 kernel audit — BUILD / DEFER / DROP](https://posttrainllm.com/docs/research/wave_2_5_kernel_audit/) — [Markdown](https://posttrainllm.com/docs/research/wave_2_5_kernel_audit.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Wave 4 landscape — TML / Apple FM / code agents / Indic](https://posttrainllm.com/docs/research/wave_4_landscape/) — [Markdown](https://posttrainllm.com/docs/research/wave_4_landscape.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Single-machine roadmap — moved](https://posttrainllm.com/docs/roadmap/) — [Markdown](https://posttrainllm.com/docs/roadmap.md) — Consolidated into docs/PLAN.md. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [Roadmap blockers — moved | PostTrainLLM reference](https://posttrainllm.com/docs/roadmap/blockers/) — [Markdown](https://posttrainllm.com/docs/roadmap/blockers.md) — Consolidated into docs/PLAN.md §2. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [Roadmap orthogonal categories — moved](https://posttrainllm.com/docs/roadmap/categories/) — [Markdown](https://posttrainllm.com/docs/roadmap/categories.md) — Consolidated into docs/PLAN.md. Several markers in the old version were stale; the merged doc is verified against code.
- [Roadmap — open-source datasets](https://posttrainllm.com/docs/roadmap/datasets/) — [Markdown](https://posttrainllm.com/docs/roadmap/datasets.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Roadmap honest summary — moved](https://posttrainllm.com/docs/roadmap/honest_summary/) — [Markdown](https://posttrainllm.com/docs/roadmap/honest_summary.md) — Consolidated into docs/PLAN.md. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [North Star — refined (2026-05-31)](https://posttrainllm.com/docs/roadmap/north_star_refined/) — [Markdown](https://posttrainllm.com/docs/roadmap/north_star_refined.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Roadmap phased plan — moved | PostTrainLLM reference](https://posttrainllm.com/docs/roadmap/phased_plan/) — [Markdown](https://posttrainllm.com/docs/roadmap/phased_plan.md) — Consolidated into docs/PLAN.md. Most phases shipped; remainder in §3. This preserved page points to the current PostTrainLLM roadmap and canonical project…
- [Roadmap recommended order — moved](https://posttrainllm.com/docs/roadmap/recommended_order/) — [Markdown](https://posttrainllm.com/docs/roadmap/recommended_order.md) — Consolidated into docs/PLAN.md §3. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [Roadmap Tier 1 — moved | PostTrainLLM reference](https://posttrainllm.com/docs/roadmap/tier1/) — [Markdown](https://posttrainllm.com/docs/roadmap/tier1.md) — Consolidated into docs/PLAN.md. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [Roadmap Tier 2 — moved | PostTrainLLM reference](https://posttrainllm.com/docs/roadmap/tier2/) — [Markdown](https://posttrainllm.com/docs/roadmap/tier2.md) — Consolidated into docs/PLAN.md. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [Roadmap Tier 3 — moved | PostTrainLLM reference](https://posttrainllm.com/docs/roadmap/tier3/) — [Markdown](https://posttrainllm.com/docs/roadmap/tier3.md) — Consolidated into docs/PLAN.md. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [Roadmap Tier 4 (skip) — moved | PostTrainLLM reference](https://posttrainllm.com/docs/roadmap/tier4_skip/) — [Markdown](https://posttrainllm.com/docs/roadmap/tier4_skip.md) — Consolidated into docs/PLAN.md §2. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [Roadmap Tier 5 (2026 frontier) — moved](https://posttrainllm.com/docs/roadmap/tier5_frontier_2026/) — [Markdown](https://posttrainllm.com/docs/roadmap/tier5_frontier_2026.md) — Consolidated into docs/PLAN.md §3 Tier 5. This preserved page points to the current PostTrainLLM roadmap and canonical project status.
- [SEO depth plan — flagship | PostTrainLLM reference](https://posttrainllm.com/docs/seo-depth-plan/) — [Markdown](https://posttrainllm.com/docs/seo-depth-plan.md) — Mac-local post-training authority plan for posttrainllm.com: artifact datasheets, train-on-Mac runbooks, benchmark leaderboards.
- [SEO sprint — PostTrainLLM technical documentation](https://posttrainllm.com/docs/seo-sprint/) — [Markdown](https://posttrainllm.com/docs/seo-sprint.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session — eval-first prep before the 2-day training window](https://posttrainllm.com/docs/sessions/06-05-eval-first/) — [Markdown](https://posttrainllm.com/docs/sessions/06-05-eval-first.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Strategy session — what posttrainllm actually is, what's…](https://posttrainllm.com/docs/sessions/06-06-mac-specialist-platform/) — [Markdown](https://posttrainllm.com/docs/sessions/06-06-mac-specialist-platform.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Specialist Embedder PRD Blocker](https://posttrainllm.com/docs/sessions/06-06-specialist-embedder-blocker/) — [Markdown](https://posttrainllm.com/docs/sessions/06-06-specialist-embedder-blocker.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [vllm-mlx Investigation | PostTrainLLM reference](https://posttrainllm.com/docs/sessions/06-06-vllm-mlx-investigation/) — [Markdown](https://posttrainllm.com/docs/sessions/06-06-vllm-mlx-investigation.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session — 2026-06-08, the factory-validation night](https://posttrainllm.com/docs/sessions/06-08-factory-validation-night/) — [Markdown](https://posttrainllm.com/docs/sessions/06-08-factory-validation-night.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Strategy session — the market landscape and the Mac-first…](https://posttrainllm.com/docs/sessions/06-13-market-landscape-mac-first/) — [Markdown](https://posttrainllm.com/docs/sessions/06-13-market-landscape-mac-first.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session — from frontier-parity distillation to the…](https://posttrainllm.com/docs/sessions/06-16-distill-to-self-improvement/) — [Markdown](https://posttrainllm.com/docs/sessions/06-16-distill-to-self-improvement.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Step-back: what we have, the wall, and the ROI menu](https://posttrainllm.com/docs/sessions/06-17-stepback-inventory-roi/) — [Markdown](https://posttrainllm.com/docs/sessions/06-17-stepback-inventory-roi.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Vercel Eve and the local-agent wedge](https://posttrainllm.com/docs/sessions/06-17-vercel-eve-and-local-agents/) — [Markdown](https://posttrainllm.com/docs/sessions/06-17-vercel-eve-and-local-agents.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Drilldown — the experiments left untried, run to bedrock](https://posttrainllm.com/docs/sessions/DRILLDOWN/) — [Markdown](https://posttrainllm.com/docs/sessions/DRILLDOWN.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM — retrospective and forward plan](https://posttrainllm.com/docs/sessions/RETROSPECTIVE/) — [Markdown](https://posttrainllm.com/docs/sessions/RETROSPECTIVE.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pace handoff — posttrainllm deliverables status (2026-06-10)](https://posttrainllm.com/docs/sessions/pace-handoff-2026-06-10/) — [Markdown](https://posttrainllm.com/docs/sessions/pace-handoff-2026-06-10.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Planner lock - 2026-06-19 | PostTrainLLM reference](https://posttrainllm.com/docs/sessions/planner-lock-2026-06-19/) — [Markdown](https://posttrainllm.com/docs/sessions/planner-lock-2026-06-19.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM — Working Session QandA Log](https://posttrainllm.com/docs/sessions/qa_log/) — [Markdown](https://posttrainllm.com/docs/sessions/qa_log.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Session retrospective — 2026-05-31](https://posttrainllm.com/docs/sessions/session_2026_05_31/) — [Markdown](https://posttrainllm.com/docs/sessions/session_2026_05_31.md) — Multi-hour session that took posttrainllm from "Wave 2.5 stalled" to "every CLI verified + first specialist trained end-to-end +…
- [Session Retrospective: posttrainllm, the Long Working…](https://posttrainllm.com/docs/sessions/session_retrospective/) — [Markdown](https://posttrainllm.com/docs/sessions/session_retrospective.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [First specialist run — toolcall-v1 findings](https://posttrainllm.com/docs/sessions/specialist_v1_findings/) — [Markdown](https://posttrainllm.com/docs/sessions/specialist_v1_findings.md) — End-to-end first specialist training on M5 Pro. What worked, what didn't, what was unblocked, and what to try next.
- [v11 baselines — committed 2026-06-09](https://posttrainllm.com/docs/sessions/v11-baselines-2026-06-09/) — [Markdown](https://posttrainllm.com/docs/sessions/v11-baselines-2026-06-09.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [WWDC 2026 (June 8–9) — impact on posttrainllm + Pace](https://posttrainllm.com/docs/sessions/wwdc-2026-impact/) — [Markdown](https://posttrainllm.com/docs/sessions/wwdc-2026-impact.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [A1 — first tool-calling specialist](https://posttrainllm.com/docs/specialists/a1-tool-caller/) — [Markdown](https://posttrainllm.com/docs/specialists/a1-tool-caller.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B1 SQL POC | PostTrainLLM reference](https://posttrainllm.com/docs/specialists/b1-sql-poc/) — [Markdown](https://posttrainllm.com/docs/specialists/b1-sql-poc.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [B1 — second specialist: text-to-SQL](https://posttrainllm.com/docs/specialists/b1-sql/) — [Markdown](https://posttrainllm.com/docs/specialists/b1-sql.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Technique Registry | PostTrainLLM reference](https://posttrainllm.com/docs/techniques/README/) — [Markdown](https://posttrainllm.com/docs/techniques/README.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Audit 2026 Technique Inventory](https://posttrainllm.com/docs/techniques/audit-inventory/) — [Markdown](https://posttrainllm.com/docs/techniques/audit-inventory.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Constrained generation: JSON-mode for posttrainllm](https://posttrainllm.com/docs/techniques/constrained_generation/) — [Markdown](https://posttrainllm.com/docs/techniques/constrained_generation.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Knowledge distillation — making a tiny model punch above…](https://posttrainllm.com/docs/techniques/distillation/) — [Markdown](https://posttrainllm.com/docs/techniques/distillation.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Evolution Strategies — gradient-free training](https://posttrainllm.com/docs/techniques/evolution_strategies/) — [Markdown](https://posttrainllm.com/docs/techniques/evolution_strategies.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [GaLore + training-stability bells (Tier 2)](https://posttrainllm.com/docs/techniques/galore_and_stability/) — [Markdown](https://posttrainllm.com/docs/techniques/galore_and_stability.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Interpretability tools — what is the model thinking?](https://posttrainllm.com/docs/techniques/interpretability/) — [Markdown](https://posttrainllm.com/docs/techniques/interpretability.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [LoRA guide — fine-tuning the tiny model](https://posttrainllm.com/docs/techniques/lora_guide/) — [Markdown](https://posttrainllm.com/docs/techniques/lora_guide.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Method vs Recipe | PostTrainLLM reference](https://posttrainllm.com/docs/techniques/method-vs-recipe/) — [Markdown](https://posttrainllm.com/docs/techniques/method-vs-recipe.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Mixture-of-Experts — more capacity per byte of weight](https://posttrainllm.com/docs/techniques/moe/) — [Markdown](https://posttrainllm.com/docs/techniques/moe.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Multi-Token Prediction — better training signal per step](https://posttrainllm.com/docs/techniques/mtp/) — [Markdown](https://posttrainllm.com/docs/techniques/mtp.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Needle 2 baseline review | PostTrainLLM reference](https://posttrainllm.com/docs/techniques/needle2-baseline-review/) — [Markdown](https://posttrainllm.com/docs/techniques/needle2-baseline-review.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Needle 45M successor factorial](https://posttrainllm.com/docs/techniques/needle2-successor-factorial/) — [Markdown](https://posttrainllm.com/docs/techniques/needle2-successor-factorial.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Optimizers | PostTrainLLM reference](https://posttrainllm.com/docs/techniques/optimizers/) — [Markdown](https://posttrainllm.com/docs/techniques/optimizers.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [parakeet.wgsl browser smoke | PostTrainLLM reference](https://posttrainllm.com/docs/techniques/parakeet-wgsl-browser-smoke/) — [Markdown](https://posttrainllm.com/docs/techniques/parakeet-wgsl-browser-smoke.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PEFT Variants in posttrainllm technical documentation](https://posttrainllm.com/docs/techniques/peft_variants/) — [Markdown](https://posttrainllm.com/docs/techniques/peft_variants.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Numerics and precision — the gate framework](https://posttrainllm.com/docs/techniques/precision/) — [Markdown](https://posttrainllm.com/docs/techniques/precision.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pruning — magnitude masks and structured surgery](https://posttrainllm.com/docs/techniques/pruning/) — [Markdown](https://posttrainllm.com/docs/techniques/pruning.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Quantization Expansion | PostTrainLLM reference](https://posttrainllm.com/docs/techniques/quantization_expansion/) — [Markdown](https://posttrainllm.com/docs/techniques/quantization_expansion.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Speculative-decoding heads: Medusa + EAGLE-2](https://posttrainllm.com/docs/techniques/speculative_heads/) — [Markdown](https://posttrainllm.com/docs/techniques/speculative_heads.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [SQL Technique Lineage — Closed](https://posttrainllm.com/docs/techniques/sql-technique-backlog/) — [Markdown](https://posttrainllm.com/docs/techniques/sql-technique-backlog.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [StreamingLLM + KIVI — long-context decode without growing…](https://posttrainllm.com/docs/techniques/streaming_llm_kivi/) — [Markdown](https://posttrainllm.com/docs/techniques/streaming_llm_kivi.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [TrainLoop Teardown | PostTrainLLM reference](https://posttrainllm.com/docs/techniques/trainloop-teardown/) — [Markdown](https://posttrainllm.com/docs/techniques/trainloop-teardown.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Tool-call extractor (mini-router)](https://posttrainllm.com/docs/tool_call_extractor/) — [Markdown](https://posttrainllm.com/docs/tool_call_extractor.md) — Design + training recipe for posttrainllm's tiny encoder model that picks which tool a user query needs, before the full LM forward pass.
- [The three phases of training — pretrain, SFT, DPO](https://posttrainllm.com/docs/training/) — [Markdown](https://posttrainllm.com/docs/training.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [DPO training: direct preference optimization for LLMs](https://posttrainllm.com/docs/training/dpo/) — [Markdown](https://posttrainllm.com/docs/training/dpo.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM — nightly training queue](https://posttrainllm.com/docs/training/nightly/) — [Markdown](https://posttrainllm.com/docs/training/nightly.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Pretraining | PostTrainLLM reference](https://posttrainllm.com/docs/training/pretrain/) — [Markdown](https://posttrainllm.com/docs/training/pretrain.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Supervised fine-tuning (SFT) | PostTrainLLM reference](https://posttrainllm.com/docs/training/sft/) — [Markdown](https://posttrainllm.com/docs/training/sft.md) — Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.