# PostTrainLLM

> A Mac-local LLM specialist factory: post-training, evidence-gated packaging, MLX runtime work, and a WebGPU learning 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): 128 resolved attempts — 5 worked, 55 worked with caveat, 68 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): Bounded full public page content

## 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.


<!-- Keep in sync with robots.txt and ai.txt. -->

## Public pages

- [PostTrainLLM — Build and measure specialist LLMs on a Mac](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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.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.md): Verified product improvements shipped by the posttrainllm Mac-local specialist factory.
- [posttrainllm — devlog](https://posttrainllm.com/devlog.md): Notes from building posttrainllm — kernel measurements, honest negative results, decisions made while AI-pairing.
- [PostTrainLLM docs technical documentation](https://posttrainllm.com/docs.md): Mac-local LLM specialist factory — training, inference, evals, systems notes, and learning paths.
- [Citations | PostTrainLLM reference](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.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.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.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.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.md): Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [PostTrainLLM Docs technical documentation](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.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.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.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.md): Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.
- [Milestones | PostTrainLLM reference](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.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.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.md): Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.

Showing 50 of 595 pages; the catalog lists all public routes.

## Machine-readable

- [Full content](https://posttrainllm.com/llms-full.txt): Bounded public page content
- [Markdown catalog](https://posttrainllm.com/api/md/_catalog): All registered public routes
- [Agent catalog](https://posttrainllm.com/api/ai): Existing discovery contract
- [AI usage policy](https://posttrainllm.com/ai.txt): Permissions and attribution
- [Sitemap](https://posttrainllm.com/sitemap.xml): Canonical HTML pages
- [OpenAPI](https://posttrainllm.com/openapi.json): Public API contract