hf-model-card-research

HF Model Card Research

Extract structured metadata — downloads, likes, benchmark claims, file sizes, author statements — from HuggingFace model cards. Used when the user asks you to "check these models on HF", "pull benchmarks for these variants", or "compare what authors claim."

When to Use

• User provides a list of authors and model families (e.g. "pull metadata for these Qwen3.6 variants")
• You need to compare benchmark claims across multiple fine-tunes of the same base model
• You need up-to-date download counts, likes, or file sizes for model variants
• Verifying what a model card claims vs what other sources say
• Surveying the ecosystem around a new base model release

Workflow

1. Discover model repo names

Search HF for each author's models using site:huggingface.co queries:

site:huggingface.co <author> <model-family> <variant-keyword>

Try variant-specific keywords: the author's handle, the model name, key terms like "uncensored", "abliterated", "Opus", "NVFP4", "MTP", etc.

For community quantizers (mradermacher, unsloth, byteshape), search with: site:huggingface.co <author> Qwen3.6-27B.

If search results are sparse, try the author's HF profile page directly: https://huggingface.co/<author>/models — use web_extract on that.

2. Pull structured stats via HF API

The HF API endpoint returns the metadata you need. Pull JSON, don't scrape the web view for stats:

https://huggingface.co/api/models/{owner}/{repo}

Returns JSON with:

• downloads — download count
• likes — like count
• pipeline_tag — model type (text-generation, image-text-to-text)
• tags, cardData, config, gguf, safetensors, siblings (file list), createdAt/lastModified
• spaces (linked HF Spaces)
• model-index (evaluation results when present)

Pitfall — delegated subagents hit terminal guardrails on HF pulls. Three dispatched subagents (each on different HF research tasks, all variants) hit same_tool_failure_halt after 4 repeated terminal retries on failed HF calls. The failure mode is predictable: subagents loop on the same failing terminal (curl to HF API) because they think "retry the same call until it works." The per-turn tool guardrail halts them after 4 non-progressing attempts with zero data produced. This happened across three separate delegations in the same session.