Development-time pipeline that builds, evaluates, and packages FreshData's small local model artifacts (fd-col-encoder-v1, fd-intent-v1, calib-v1). This package is not part of the runtime: it lives outside src/freshdata, is excluded from the wheel (see pyproject.toml and .github/workflows/wheel-size.yml), and the runtime never imports it.

FreshData's runtime stays deterministic, offline, model-free by default, and free of any LLM or cloud call — nothing here changes that contract. See docs/developer-training-pipeline.md for the full guide and COMPLIANCE.md for the teacher-model governance rules.

training/
  seed/          seed corpus registry, licenses, synthetic PII-shaped generators
  corruptors/    ~40 deterministic, labeled corruptors (representation/semantic/context/scale)
  teacher/       compliance-gated development-time teacher harness (schemas, cache, review)
  datasets/      seed/corrupted dataset builders, splits, license+label validators
  distill/       role-head / intent-head / optional encoder training, ONNX export, quantize, package
  calibration/   feature extraction, isotonic fit, table export, eval
  eval/          human-verified release-gating label store
  model_cards/   the three required model cards
  cache/         teacher prompt/response audit cache (gitignored content, tracked directory)

make training-dev-artifacts       # synthetic-only, offline, seconds
make training-release-artifacts   # full gated pipeline (compliance -> package -> CleanBench)
python -m benchmarks.cleanbench --tracks T1,T2,T3,T4,T5 --report site

Individual stages: make training-seed, training-corrupt, training-teacher-labels, training-distill, training-eval, training-package-artifacts (see training/Makefile).

Corruption metadata is the preferred ground truth; teacher output is training data only and never gates a release on its own — release-gating eval labels must be 100% human-verified (training/eval/human_verified.py).