Skip to main content

Publish a Model Update

Publishing a model update is similar to publishing a new model. The difference is that the model weights are the result of pretraining, fine-tuning, or aggregation of model deltas, rather than the creation of randomized model weights.

To publish a model update:

  1. Create Safetensors weights from your traind model checkpoint:
goodwoolf save <vendor>/<model> \
--load-path $GOODWOOLF_MODELS/<vendor>/<model>/chpt.pt \
  1. Upload your initial model weights to a public URL (e.g., a Hugging Face model):
  • $GOODWOOLF_MODELS/<vendor>/<model>
  1. Create a PR for the Training Ledger repo with the ledger file:
  • <gw_ledger_repo>/<vendor>/<model>/training.ledger.json

Add a new entry to the model ledger:

goodwoolf ledger <vendor>/<model> --create pretrained

Update the training.ledger.json in (3) and set the new entry's url to the URL in (1) where the updated model weights are uploaded. Set the operator to pretrained, finetuned or aggregated, depending on how the weights were created. Set the inputs to the base training model or a list of aggregated models, each referenced by its entry ID. Check the entry:

goodwoolf ledger <vendor>/<model> --check

The remaining steps to publish the model updates PR are the same as for the new model.