Aads Training Lifecycle

Maintain AADS-ULoRA phase1, phase2, and phase3 training behavior plus adapter lifecycle invariants across local and Colab flows. Use for training logic, freezing rules, retention/OOD threshold behavior, and phase3 runtime or component updates.

majiayu000 13cf88e 2 files · 2.2 KB Updated 567 repo stars

File contents

AADS Training Lifecycle

Workflow

  1. Identify phase scope.
  • Determine whether change targets Phase 1, 2, 3, or cross-phase behavior.
  • Load references/training-invariants.md.
  1. Preserve lifecycle invariants.
  • Preserve expected freezing and adaptation semantics.
  • Preserve OOD-threshold path compatibility with adapters.
  • Preserve phase3 runtime/component contract behavior.
  1. Enforce local/Colab parity.
  • Review paired local and colab_* trainer variants when behavior changes.
  • Include notebook/script entrypoint checks when user-facing training flow changes.
  1. Run phase-aligned validation.
  • Run targeted training unit tests.
  • Run Colab smoke training checks when applicable.
  • Add integration checks only when required.
  1. Emit training delta summary.
  • Report phase-wise behavior changes and invariants status.

Output Contract

Return sections in this order:

  1. Phase scope
  2. Files changed
  3. Invariants preserved/changed
  4. Validation commands and results
  5. Risks and follow-up checks
  6. Local-Colab parity notes

References

  • Read references/training-invariants.md before implementing.

majiayu000/claude-skill-registry-data/tree/main/ai-ml/aads-training-lifecycle commit 13cf88e384

Frequently asked questions

npx skillmds add majiayu000/aads-training-lifecycle