CLI reference

goldilocks-ml train    seal | validate | run
goldilocks-ml publish  validate | checksum | upload

Run them from a clone with uv run. Every command exits non-zero and shows the real error when something fails.

There is no inference command. Predicting from a published model is Goldilocks Core's job; this package gives it a library.

train seal

uv run goldilocks-ml train seal DATASET \
  --record-id RECORD --version VERSION \
  --target TARGET --target-contract CONTRACT \
  --target-definition DEFINITION [--target-units UNITS]

Writes manifest.json with the snapshot's identity, target definition, and the size and SHA-256 of every file. Offline.

train validate

uv run goldilocks-ml train validate PROTOCOL --dataset DATASET

Checks the protocol against the snapshot, derives the split, and reports the per-split counts. Trains nothing. Offline.

train run

uv run goldilocks-ml train run PROTOCOL --dataset DATASET --output OUTPUT \
  [--splits SPLITS] [--artifact-directory DIR] [--overwrite]

Runs every validate check, then trains and writes a run bundle.

--splits replays an existing splits.csv. --overwrite only replaces a directory an earlier run created; it refuses ordinary directories.

publish checksum

uv run goldilocks-ml publish checksum PATH

Prints one manifest entry — name, size, digest — ready to paste. Offline.

publish validate

uv run goldilocks-ml publish validate DEPOSITION --artifact-directory DIR

Checks the metadata, the model card, and every artifact's size and digest. Offline.

publish upload

uv run goldilocks-ml publish upload DEPOSITION --artifact-directory DIR \
  --token-file TOKEN_FILE --confirm-upload

Validates again, creates a PSDI draft, uploads the files, and prints the draft id. It never submits for review — do that on the website.

If a step fails partway, the partial draft is deleted. If that cleanup also fails, the error names the draft id so you can remove it yourself.