菜单
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.
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 Use BEFORE writing any ML training, fine-tuning, or inference code — your knowledge of library APIs is outdated and example scripts show the current patterns.
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.
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-datasets |
| description | 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. |
Backed by scripts/inspect_dataset.py.
Always call this before training. Training fails with KeyError if columns don't match the method. Specifically:
messages, text, or prompt + completion.prompt, chosen, rejected.prompt.Also use to:
# Full inspection — auto-detects config and split
python scripts/inspect_dataset.py --dataset stanfordnlp/imdb
# Pick a specific split + more sample rows
python scripts/inspect_dataset.py --dataset HuggingFaceH4/ultrachat_200k --split train_sft --sample-rows 5
# Private / gated dataset — HF_TOKEN must be set
HF_TOKEN=hf_... python scripts/inspect_dataset.py --dataset org/gated-dataset
Output is markdown with sections:
--sample-rows rows, long values truncated at 150 chars.messages is present) — roles seen, which keys exist (role, content, tool_calls, tool_call_id, name, function_call), tool-call/result presence, and an example message structure.config/split.--dataset (required) — org/name format.--config, --split — auto-detected if omitted.--sample-rows N — default 3, cap 10.--json — wrap output in {"ok": bool, "output": "..."}.HF_TOKEN — only needed for private / gated datasets.hf-papers → find_datasets surfaces datasets linked to a paper. Pipe the top hit into this skill.hf-repos → for deeper repo-level file access (e.g. read a custom dataset_infos.json).