Skip to content

ROLLING_TRAINING.md

Latest commit

 

History

History
36 lines (27 loc) · 1.25 KB

ROLLING_TRAINING.md

File metadata and controls

36 lines (27 loc) · 1.25 KB
Raw

Rolling training

layercake/rolling/ is the training-control substrate for future LayerCake iteration. It is designed for small, reversible semantic updates instead of opaque checkpoint overwrites.

Core contract:

  • train from a declared TrainingRubric;
  • commit every candidate as a content-addressed ModelCommit;
  • run semantic gates before promotion;
  • preserve failed commits;
  • roll back trainable modules to the last passing parent when gates fail;
  • keep ABI, input-interface, and byte-patch hashes stable unless a rubric explicitly targets an ABI migration.

Smoke demo:

python scripts/demo_rolling_training.py --smoke

Benchmarks:

python scripts/benchmark_rolling_training.py
python scripts/benchmark_rollback_cost.py
python scripts/benchmark_cherrypick_transfer.py

The current demo is intentionally tiny. It proves the mechanics: pass, fail, rollback, safe follow-up, semantic certificate emission, capability ledger entry, and module cherry-pick under compatible ABI hashes. It does not prove language-model quality.

The next layer is preview-guided rolling training. It inserts a non-destructive preview and compiled syllabus before staged training. See PREVIEW_GUIDED_TRAINING.md.