菜单
Explore and fetch HuggingFace library documentation (transformers, trl, datasets, peft, accelerate, trackio, vllm, optimum, courses, etc.). Use when you need current trainer arguments, config parameters, or API reference that your internal knowledge may have wrong.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Alternative paper search backends — OpenAlex (~250M works, no auth, strong PDF extraction) and CORE (~49M full-text, BODY search). Use when hf-papers can't find a paper, when Semantic Scholar is rate-limited, when you need PDFs of older/long-tail papers, or when you need to search the actual body text of papers (not just titles/abstracts).
Discover GitHub repositories, find example scripts within them, and read source files with line-range support and automatic Jupyter notebook conversion. Use BEFORE writing any ML training, fine-tuning, or inference code — your knowledge of library APIs is outdated and example scripts show the current patterns.
Inspect a HuggingFace dataset in one call — status, configs/splits, schema, sample rows, parquet file listing, and detailed messages-column analysis (roles, keys, tool-calls). REQUIRED before submitting any training job to verify column format matches the training method.
Submit and manage HuggingFace compute jobs — run training / fine-tuning / inference / batch work on HF cloud infrastructure in Python or Docker mode. Supports one-shot and recurring (cron-style) jobs.
Discover ML research papers, read their contents by section, trace citations, find linked datasets/models — combining HuggingFace Papers, arXiv, and Semantic Scholar. Use whenever you need to research a method, find training recipes with published results, crawl a citation graph, or locate the datasets a paper used. This is your starting point for any literature-first workflow.
File and git-like operations on HuggingFace repositories (model / dataset / space). Covers list / read / upload / delete of files, branches, tags, pull requests, and repo management (create / update). Use to inspect repos, ship artifacts, and propose changes via PRs.
| name | hf-docs |
| description | Explore and fetch HuggingFace library documentation (transformers, trl, datasets, peft, accelerate, trackio, vllm, optimum, courses, etc.). Use when you need current trainer arguments, config parameters, or API reference that your internal knowledge may have wrong. |
Backed by scripts/hf_docs.py. Two subcommands:
explore — list pages in a doc set, optionally filtered by a keyword query.fetch — fetch the full markdown content of a specific page URL.unexpected keyword argument error.Pass any of these to --endpoint:
transformers, datasets, tokenizers, accelerate, peft, trl, diffusers, trackio, safetensors.text-generation-inference, text-embeddings-inference, inference-endpoints, vllm, lighteval.hub, huggingface_hub, hub-docs, datasets-server.optimum → optimum + optimum-habana + optimum-neuron + optimum-intel + optimum-executorch + optimum-tpu.courses → llm-course + robotics-course + mcp-course + smol-course + agents-course + deep-rl-course + computer-vision-course + audio-course + ml-games-course + diffusion-course + ml-for-3d-course + cookbook.# Find the page(s) matching your question
python scripts/hf_docs.py explore --endpoint trl --query sft
python scripts/hf_docs.py explore --endpoint transformers --query 'flash attention'
# Fetch the full markdown of the most relevant result
python scripts/hf_docs.py fetch --url https://huggingface.co/docs/trl/sft_trainer
whoosh is installed (via pip install -e '.[docs]'), search uses a proper stemming full-text index.HF_TOKEN — required.explore:
--endpoint (required), --query (optional), --max-results N (default 20, cap 50).fetch:
--url (required) — can be the .md URL or the plain doc URL (.md is auto-appended).Both accept --json.
github-code — use together: docs tell you what parameters exist, example scripts show them in use.hf-papers — papers describe what to do; docs describe how the current API supports it.