What it does
Inspects a model checkpoint and an associated eval log, then produces a structured model card following the Hugging Face / Mitchell et al. template. The card covers intended use, training data summary, eval results, known limitations, and ethical considerations.
Why it’s useful
Model cards are increasingly required by procurement, governance, and downstream consumers. Hand-writing them is tedious and error-prone. This tool extracts what it can automatically and prompts you for the parts that require human judgment.
How to use it
- Upload or link a checkpoint (HF Hub, local path, or remote URL).
- Attach an eval log in standard JSON format.
- Fill in the human-judgment sections (intended use, limitations).
- Export as Markdown, HTML, or a Hugging Face Hub model card.
Limitations
- Beta: schema may shift as Hugging Face’s model-card spec evolves.
- Eval log parsing currently supports lm-eval-harness and OpenCompass formats. Others welcome via PR.
- Doesn’t generate model cards for closed/proprietary models — provenance must be verifiable.