ml-repro

Files (1)

• SKILL.md

SKILL.md (first lines)

When writing or reviewing ML code, verify:

0. Python >=3.11, recorded in project config; the actual interpreter and notebook
   kernel verified (details in `repo-bootstrap`).
1. Seeds fixed and logged (framework, numpy, dataloaders).
2. Experiment configuration versioned in files, not
   hardcoded or only in notebooks.
3. Identifiable data version (hash, snapshot, or immutable
   reference). Never paths to data that can change without trace.
4. Metrics and artifacts logged in the project tracker (see
   repo AGENTS.md); if the repo does not define a tracker, ask before
   introducing a new one.
5. Train/val/test separation without data leakage; document the split.
6. The experiment can be rerun with one documented command.

For notebooks: reusable logic extracted into modules; the notebook
only orchestrates and visualizes.

From G4rb4sh/MLE-setup, MIT.

More from G4rb4sh/MLE-setup