Hugging Face Hub
Feynman includes read-only Hugging Face Hub tools for grounding ML recipes and replication plans. They help the researcher verify whether a dataset or repo actually exposes the files, splits, schema, and card metadata needed for implementation.
These tools are grounded in Hugging Face’s public Hub API endpoint docs. Authentication follows the HF_TOKEN environment variable documented by huggingface_hub; Feynman also accepts HUGGINGFACE_HUB_TOKEN for compatibility with existing shells.
Authentication
Public Hub resources work without configuration. For private or gated resources, set an access token in your shell before launching Feynman:
export HF_TOKEN=hf_...Feynman also checks HUGGINGFACE_HUB_TOKEN. The tools send whichever token is present with Hub requests.
Tools
The researcher agent can use these tools automatically:
| Tool | Purpose |
|---|---|
hf_dataset_info | Inspect dataset metadata, tags, access status, card data, features, splits, downloads, likes, and sibling files |
hf_repo_files | List files in a model, dataset, or Space repository before reading anything large |
hf_repo_read_file | Read small text files from Hub repos, such as README.md, configs, examples, and scripts |
The file reader truncates output by default and is intended for text files only. It is not a weight downloader or dataset bulk reader. It refuses obvious model weight files, archives, and dataset shards such as .safetensors, .bin, .gguf, .parquet, .zip, and .tar before download.
Where it is used
The /recipe workflow uses Hugging Face Hub inspection to validate dataset availability and schema before recommending a training recipe. The /replicate workflow uses the same checks for ML-heavy papers and benchmark claims.
For example, a recipe using a chat SFT dataset should verify that the dataset has a messages, text, or prompt/completion-style schema before calling it usable. If Feynman cannot check that schema, the final artifact should mark the dataset as unverified or blocked.
Boundaries
These tools are read-only. They do not create repos, upload files, start jobs, or manage private data. If an experiment needs execution on Hugging Face infrastructure, Feynman should record that as a follow-up implementation decision rather than silently attempting it.