Summary

• Adds hf models card <model_id> and hf datasets card <dataset_id> commands that print the repo card (README) to stdout
• Three output modes: full card (default), --metadata (just the YAML frontmatter as JSON), --text (just the markdown body)
• --metadata and --text are mutually exclusive

Motivation

hf models info gives you structured Hub metadata (downloads, tags, pipeline_tag, siblings, etc.) but not the human-authored card content. The card text is where you find the stuff that info doesn't surface: usage examples with actual code, training details, known limitations, intended use cases, benchmark results with context, and architecture descriptions. Put simply — info tells you what a model is, card tells you how to use it and why.

info does include a card_data field, but it's the raw YAML string, not parsed. --metadata returns the same data as structured JSON via out.dict(), so it works with --format and is easy to pipe into jq or consume programmatically.

Agents and humans can already get card content via hf download <repo_id> README.md or curl, but that writes to a file and gives you the raw README with no way to split the YAML frontmatter from the prose. hf models card outputs directly to stdout and the --metadata/--text flags let you grab just the part you need.

For agents specifically, having a low-friction way to read model documentation helps reduce hallucination. Agents tend to default to recommending models they've memorised from training data (often outdated — e.g. still reaching for early Llama models), and fabricate usage details rather than checking the actual card. A single command that returns the real card content makes it easy for agents to look things up rather than guess. This is particularly valuable for newer models that post-date the agent's training cutoff.

For humans, it's a quick way to check a model's docs from the terminal without opening a browser — useful when comparing models or scripting.

Examples

# Full card to stdout
$ hf models card google/gemma-4-31B-it

# Just the card metadata (from the YAML frontmatter)
$ hf models card google/gemma-4-31B-it --metadata

# Card metadata as JSON
$ hf models card google/gemma-4-31B-it --metadata --format json
{"library_name": "transformers", "license": "apache-2.0", ...}

# Pretty-printed
$ hf models card google/gemma-4-31B-it --metadata --format human

# Just the text body (no YAML frontmatter)
$ hf models card google/gemma-4-31B-it --text

# Same for datasets
$ hf datasets card HuggingFaceFW/fineweb --metadata --format human

Design notes

• --metadata not --yaml — We considered --yaml (the source format) and --frontmatter (the structural term) but went with --metadata because it describes what you're extracting rather than where it lives. It also pairs cleanly with --text — both flags describe the kind of content you want. And it avoids confusion with --format, which controls output format: --metadata --format json reads clearly as "give me the metadata, formatted as JSON".
• No --revision support — RepoCard.load() doesn't currently pass revision to hf_hub_download. Could be added to RepoCard.load() in a follow-up and then wired through here.
• --format is accepted even though the default and --text modes output free-form text (where --format json produces no output, same as hf papers read). We kept it because --metadata goes through out.dict() and genuinely benefits from it (e.g. --format human for pretty-printed JSON). This follows the majority CLI pattern — hf papers read is the only command that omits --format.

🤖 Generated with Claude Code