MIT · one developer · Apple Silicon
posttrainllm turns a stock open model into a routed specialist that earns its keep — then tells you exactly where it fails. Distill, fine-tune, gate on real evals, package for MLX, and open the model up with interpretability. 100+ CLI subcommands, one Apple Silicon machine, no cloud.
frozen5,675 rowsdpo · 200 stepsexec 0.920tgla / mlxretry-data
Every run assembles a schema-valid folder — config, dataset, eval, decision,
report — that factory-run validates. The one proof still missing
is a run whose decision is ship.
A specialist beats a generalist on its target and routes away when it shouldn't answer. Every result here ships with the regression it costs.
Distilled to 100% on the multi-turn file-ops hard gate (BFCL), up from 58%. The honest cost: out-of-domain breadth dropped 60% → 42% — real catastrophic forgetting. So it ships routed, never as a general planner.
Two adapters behind a router: public schema-SQL and local execution-SQL. Two reference-anchored DPO retries cured a policy collapse and pushed execution to 0.920 — but output hygiene is a base-model prior a small adapter can't strip. Decision: retry-data.
Every number below was recorded on a single M5 Pro / 48 GB. No estimates.
Not a wrapper. Every capability below is a real subcommand with tests behind it.
Pretrain, SFT, DPO / SimPO / KTO / ORPO, distillation, ES. Full PEFT bundle — LoRA, LoRA+, DoRA, VeRA, LoftQ, AdaLoRA, PISSA. WSD schedules, spike recovery, z-loss.
BFCL, τ-bench, lm-eval (MLX adapter), HumanEval + sandbox, SQL execution, router, MILU, MTEB. Frozen baselines, slice metrics, exit-non-zero gates.
OpenAI- and Ollama-compatible on one socket. Agent loop, tool dispatch, FSM-constrained JSON, persistent KV cache, speculative decoding, optional cloud escalation.
Export to MLX, safetensors, CoreML. Quantize (GGUF / AWQ / GPTQ / HQQ), prune, merge, bake-LoRA — DoRA magnitudes included. Specialist model cards.
SAE features, ROME, MEMIT, tuned / logit lens, activation patching, linear probes, attention heatmaps. Know where the model decides, not just what it answers.
The same model trains in a browser tab via hand-written WebGPU kernels — Memory64, FlashAttention-2, blocked matmul. A from-scratch learning track, honest negative results and all.
A factory is a loop, not a demo. Failed runs become the next recipe — and all of it is written down, with the decision it forced.
Ref-free SimPO collapsed the SQL policy. Execution 0.860 → 0.080, degenerate fence-spam.
→ retry-training preference tuning needs a reference anchor.
Reference-anchored DPO fixed the collapse. Execution recovered to 0.900, then 0.920 at higher pressure — no regression.
→ retry-data execution was never the problem; the Answer: wrapper is a base-model prior.
Diagnosis corrected in the open. The 108 SFT targets were already bare SQL — so hygiene needs a generation-strength fix, not more DPO.
→ next: constrained SELECT-prefix decoding.
A ground-up curriculum and per-technique explainers — each tied to the paper it cites and the file where it lives.
From y = mx + b to a self-improving factory. Every subsystem in the repo maps to an anchor in the coverage map — nothing is a black box.
The headline result, the recipe, the honest tradeoffs.
interpretabilityLogit / tuned lens, patching, SAE, ROME, MEMIT.
distillationKL + NLL mix, temperature, α, rejection sampling.
evalsDrop a JSONL; compare by step, model, task.