Skip to main content

Training Ledger

GoodWoolf Training Ledger is a Git repository hosted on Hugging Face. It maintains a dedicated JSON file for each model under training, providing an organized record of individual training activities.

Each model update is represented by an entry in the file, using the following format:

{
    "id": "<your_model_hash>",
    "provider": "<your_user_name>",
    "url": "<your_model_url>",
    "weights": {
        "model.safetensors.index.json": "<sha1sum>",
        "<safetensors_filename>": "<sha1sum>",
        "<additional_safetensors_filename>": "<sha1sum>"
    },
    "inputs": [
        "<input_model_uuid>",
        "<additional_input_model_uuid>"
    ],
    "operator": "<method_used_to_create_the_model>"
}

The id entry is the SHA-1 hash of all the files listed in weights, combined.

The supported operators are:

  • randomized: Initialized a model with randomized weights (0 inputs).
  • pretrained: Pretrained a model on top of a base model (1 model input).
  • finetuned: Finetuned the model on top of a base model (1 model input).
  • aggregated: Aggregated weight deltas from multiple trained models onto a base model (1 or more inputs).