run-dig

Recover ML experiment metadata and infer lineage from existing files. No prior tracking setup required.

Requires Python 3.10+.

pip install run-dig
run-dig scan ./experiments --open

Opens a local HTML report and saves run-dig-report/index.html and report.json. Hover a graph node to highlight its links; click to see its files, metadata, and relationship evidence. Arrows show recorded or possible sources; undirected links show similarity or identical files.

Try the interactive demo — explore sample experiments in your browser, no installation needed.

Supports .pt, .pth, .ckpt, .safetensors, and model .npz files. NPZ detection uses array names and file context to distinguish models from feature caches; it does not verify model validity. Arrays are never loaded.

Use the terminal to list, inspect, or compare checkpoints:

run-dig list ./experiments --sort size --name best
run-dig inspect run-a/best.pt --root ./experiments
run-dig compare run-a/best.pt run-b/best.pt --root ./experiments

--root includes related runs; --full shows all fields and Git history for inspect and compare. Scanning is read-only. Checkpoints and scripts are never executed.

With Git installed, scans also show local commit history and verify git_commit or git.commit values recorded in run files. Current HEAD is only scan-time context; Git ancestry does not prove checkpoint inheritance. No network access is used. Pass --no-git to skip Git inspection.

Relationships and checkpoint metrics include evidence and confidence. Similarity does not prove ancestry, and confidence is not a calibrated probability. Missing or conflicting settings remain unknown; conflicting values keep their sources.