Factory Docs

Start here for active posttrainllm work.

posttrainllm’s current product loop is:

target -> data -> post-training -> eval -> package -> report

Files

../README.md — golden path through state, attempts, reviewed products, roadmap, and learning.
overview.md — what the factory is and what counts as proof.
post-training-factory.md — how data, post-training, evals, performance, packaging, and public artifacts fit together.
../attempt-ledger.md — worked, failed, regressed, inconclusive, and not-yet-tried attempts.
../external-products-reviewed.md — external products and techniques reviewed or adopted.
case-study-template.md — public artifact report shape: baseline, failed attempts, slices, trace review, performance, and blockers.
batch-posttraining.md — batch rollout, offline scoring, compact adapter update, eval, and decision loop.
lora-geometry.md — adapter effective-update diagnostics for rank and module targeting.
Autocorrect specialist lane (in progress, untrained): autocorrect-foundation.md — contract, evaluator, simulator, and manifests; autocorrect-model-shortlist.md — the measured base bake-off and selection; autocorrect-adapter-recipe.md — the frozen LoRA recipe, the encoder-decoder training path, and its load-parity evidence.
run-schema.md — local run directory contract.
run-lifecycle.md — durable phase/revision state, metadata-only operator commands, advisory discovery, and recovery.
report-card.md — portable before/after proof contract: measurement states, decision semantics, JSON + static report, publication gate.
report-card-cohort.md — the published report-card cohort, documented absences, and the mapping gaps the review found.
enforcement.md — native validation plus stricter publish-check requirements.
eval-protocol.md — baseline, regression, and ship/reject rules.
packaging.md — specialist package layout and lock metadata.
reports.md — before/after report template.
public-artifacts.md — public artifact registry, release states, and blockers.
../techniques/README — method-vs-recipe registry. Use this before selecting a post-training run so “try DPO/RLVR/LoRA” becomes a concrete recipe with data, eval, slices, and stop rule.

Rule

Use existing primitives first. Add new tooling only when it directly improves data preparation, post-training, eval, packaging, or reporting for the current factory target.

Do not treat a method name as a plan. A run is ready only when it has a recipe: target, failure mode, data, reward or labels, eval gates, slice gates, and a ship/retry/reject threshold.