| name | hf-hub |
| description | Operate the Hugging Face Hub with the `hf` CLI: auth, browse/search models, datasets, spaces and papers, download and upload files, query datasets, run training jobs, and deploy inference endpoints. Use when the user mentions hf, huggingface, or wants to use, search, or publish models, datasets, or demos on the Hub.
|
| stack | ml |
Hugging Face Hub (hf CLI)
The hf CLI is the universal entry point for the Hub: models, datasets,
spaces, buckets, papers, jobs, and endpoints. It replaced the deprecated
huggingface-cli.
Install + auth
curl -LsSf https://hf.co/cli/install.sh | bash -s
hf auth whoami
hf auth login
The rule: query help, don't memorize flags
The hf CLI changes between releases. For exact flags, always ask the CLI
first instead of guessing:
hf --help
hf models --help
hf datasets --help
hf spaces --help
hf jobs --help
Common workflows
- Search/browse:
hf models list --search "<query>", hf datasets list --search "<query>", hf spaces search "<query>", hf papers search "<query>".
- Download:
hf download <repo_id> (models/datasets/spaces).
- Upload:
hf upload <repo_id> <local_path> --commit-message "..." (single-commit uploads).
- Datasets:
hf datasets info <id>, hf datasets parquet <id>, hf datasets sql "<duckdb query>".
- Papers:
hf papers search "<query>", hf papers <arxiv-id>.
- Jobs (cloud GPU training/inference):
hf jobs run <image> <command> --flavor <t4-medium|l4x1|a100-large|...>, hf jobs logs <job-id>, hf jobs hardware.
- Inference endpoints:
hf endpoints deploy, hf endpoints list.
- Cache:
hf cache list, hf cache prune.
When to use
The user mentions hf / huggingface / a model or dataset repo id, or wants to
download a model, push a checkpoint, run a training job, or inspect a dataset.
Pitfalls
- Don't memorize the flag list; run
hf <sub> --help at execution time.
- Auth is now
hf auth ..., not the old huggingface-cli login.
- Gated/private repos need
hf auth login first, then use --token or a
logged-in session.
Adapted from huggingface/skills (Apache-2.0).
