Skip to content
posttrainllm
posttrainllm
a one-machine research lab · MITthe local research bench

build a specialist.
prove it on one Mac.

Turn a stock open model into a routed specialist on one Mac — then measure every win and regression. Distill, fine-tune, evaluate, package for MLX, and inspect the result.

run the factoryinspect specialist proof
native Mac app · Verified and ready

paired proof · research routed-ship

qwen3-4b, ReST fused.

The file-operations route improved depth, removed unexpected side effects, and ran 2.42× faster. Breadth fell on the same frozen evaluation. Retain behind the file-ops route; reject as the general successor.
stock → ReST · paired frozen eval+3 −1
  1. 1+file-ops 9/12 → 12/12
  2. 2−breadth 30/45 → 25/45
    regression — five net breadth cases lost. The depth win does not justify general promotion.
  3. 3+side effects 8 → 0
  4. 4+depth wall 360.50s → 148.91s

the thesis

reach frontier capability. understand the machine.

Be best-in-class at what one Apple Silicon machine can actually do: train, post-train, evaluate, serve, inspect. No cluster and no cloud dependency.

learn the whole space
From y = mx + b to a self-improving factory. Reproduce techniques locally or integrate the best tools, anchored to evidence and the files where they live.
show the scars
Failed runs are first-class. Every attempt keeps the decision it forced. A result carries its regression. Build everything buildable here, then package it so the lab can scale when compute arrives.

the factory loop

each surface feeds the next.

Every run ends in a schema-valid folder — config, dataset, eval, decision, report — that factory-run validates.
a bounded file-operations run
target      file operationsdata        interleaved + depth anchorpost-train  teacher-free ReSTeval        frozen depth + breadth gatespackage     fused bf16 · HF/MLXdecide      routed-ship

report-only candidate · retry-data

qwen3-0.6b, routed SQL.

Reference-anchored DPO cured a policy collapse and improved synthetic execution. Output hygiene did not improve: the small adapter could not strip the base model’s prose wrapper. This remains a report-only candidate.
frozen SQL gates · report-only+1 −1
  1. 1+execution 0.860 → 0.920
  2. 2−clean-SQL 0.000 → 0.000
    blocked — clean output still fails. Decision: retry-data; execution alone is not a ship gate.

honest scope

a factory for specialists.

Models and runs are deliberately bounded to the hardware. A specialist beats a generalist on its target and routes away when it should not answer.

what it is
A single-developer, MIT-licensed research lab. Mac-first: M-series, unified memory, MLX-Swift. An OpenAI-compatible runtime. A research-only routed ship, with missing evidence disclosed.
what it isn’t
A general assistant, multi-GPU or distributed training, or a cloud product. The ReST package is not product-default proof: runtime promotion still needs a fresh Pace-specific gate. Nothing leaves the laptop unless you ask.

run it · one Mac · from source

build. inspect. validate.

Use the native quickstart for a real Mac-local factory run. The browser playground is a bounded learning demo, not the specialist factory.
zsh — build and validate the native factory
# build the native factorygit clone https://github.com/PostTrainLLM/posttrainllm.gitcd posttrainllm/native-macxcrun swift build -c releaseBIN=.build/release/posttrainllm# inspect the current factory surface$BIN --help$BIN factory-run validate ../runs/<run-id>

attempt ledger

the result distribution, without spin.

127 resolved attempts: 5 worked cleanly, 55 worked with caveat, and 67 failed, regressed, were rejected, inconclusive, or superseded. A caveated mechanism proof is not a positive product result. Boundary evidence stays visible.

choose your entry point.

Run the native factory, inspect a bounded browser mechanism, read the evidence, or follow the learning journey.

build the specialist, then prove it.

Evidence-first field guides connect the public factory to concrete Apple Silicon workflows.