Factory Overview

TinyGPT’s active product is the specialist factory.

The factory should turn a base model plus task data into a measured specialist artifact:

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

For the fuller project/learning split and the post-training pillar map, see post-training-factory.md.

Factory Contract

A valid factory run has:

• a frozen target
• a frozen baseline
• a dataset manifest
• a training config
• an eval result before and after training
• a packaged artifact or explicit rejection
• a report with cost, latency, regressions, and decision

Anything else is research or tooling. It may be valuable, but it is not the factory proof.

Current Assets

Use these before adding new tooling:

• Data: traces-to-data, corrections-to-data, quality-filter, reasoning-classify, dedupe, download-dataset, extractor-data.
• Post-training: sft, dpo, distill, es, merge, bake-lora.
• Evals: eval-gate, eval-compare, eval-bfcl, eval-tau-bench, run-lm-eval, eval-humaneval, eval-sql, eval-router, eval-scaledown, eval-escalate.
• Packaging: export-mlx, specialist package directories under specialists/, tinygpt.project.json / lock metadata.
• Runtime checks: serve, bench, run-bench, smoke scripts in evals/.

First-Class Output

The output of the project is not just a model. It is a folder that proves what happened:

runs/<date>-<target>/
  config.json
  dataset.json
  train.log
  eval-baseline.json
  eval-candidate.json
  report.md
  artifact.json
  decision.json

runs/ is local output and is gitignored. Commit schemas, fixtures, and shipped specialist package metadata instead.

Factory Run Center

The polish/UI center should be a readout of the factory, not a separate product surface yet.

Minimum useful readout:

• runs
• target
• dataset version
• base model
• training method
• eval score
• regression score
• cost/time
• latency/RAM/tok-s
• artifact path
• decision

CLI first. UI second.