What does this skill do, and when should you use it?
This Agent Skill targets Hugging Face Dataset Viewer API workflows for inspecting metadata, browsing rows, and extracting dataset content. It covers configuration and split resolution, pagination, text search, predicate filtering, Parquet URL discovery, size totals, and column statistics. Its main workflow is read-only, but the document also includes browser and CLI procedures for creating datasets and uploading Parquet files. It additionally explains how to upload raw Claude Code, Codex, and Pi Agent session traces, with privacy guidance for potentially sensitive records.
Calls the /is-valid, /splits, /first-rows, /rows, /search, /filter, /parquet, /size, /statistics, and /croissant endpoints at https://datasets-server.huggingface.co; reads rows with offset and length pagination; searches string columns; filters rows with where and orderby parameters; discovers Parquet shard links; and reads dataset totals, column statistics, and optional Croissant metadata. It also documents creating private dataset repositories and uploading Parquet files through the Hub UI, npx @huggingface/hub, or the hf CLI.
- A data analyst needs to verify a Hugging Face dataset and resolve its available configurations and splits.
- A researcher needs to inspect dataset rows in pages rather than requesting a large result at once.
- An engineer needs to search string columns or filter and sort rows with Dataset Viewer predicates.
- A data pipeline author needs Parquet shard URLs together with dataset size or column statistics.
- A team using Claude Code, Codex, or Pi Agent needs to publish session JSONL traces to Hugging Face Datasets with privacy-conscious defaults.
How do you install this skill?
- The skill is presented as read-only but later includes dataset creation, parquet uploads, and agent-trace uploads; obtain explicit confirmation and verify the repository, visibility, paths, and token permissions before uploading.
- Agent traces may contain prompts, file paths, tool outputs, secrets, or PII; a private repository does not replace pre-upload review, redaction, and credential rotation.
- Endpoint behavior, response fields, limits, and reachability are not verified in the supplied file; validate service availability, authentication failures, rate limits, and pagination boundaries in the target environment.
- Shell / CLI
- Network access
- Local filesystem
curlHugging Face CLInpx @huggingface/hub
The repository bundles 26 skills. According to the README, for Codex, copy or symlink skills/huggingface-datasets/ into a standard Codex .agents/skills location, such as $REPO_ROOT/.agents/skills or $HOME/.agents/skills. Once present, an Agent Skills-compatible client can discover its SKILL.md. The README does not provide a standalone installation command for this skill.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/huggingface/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/huggingface-datasets" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Inspect the configurations and train split for stanfordnlp/imdb, fetch the first 100 rows, and search for records containing the specified text.
After the skill is loaded by a coding agent, give a concrete request such as: "Inspect the configurations and train split for stanfordnlp/imdb, fetch the first 100 rows, and search for records containing the specified text." A supported request can use a command such as curl "https://datasets-server.huggingface.co/rows?dataset=stanfordnlp/imdb&config=plain_text&split=train&offset=0&length=100". Gated or private datasets require Authorization: Bearer <HF_TOKEN>. For pagination, increase offset and use num_rows_total, num_rows_per_page, and partial to determine continuation.
What are this skill's strengths and limitations?
- Covers the main Dataset Viewer API endpoints for exploration, retrieval, pagination, and statistics.
- Includes copyable curl examples and a concrete pagination pattern.
- Also documents Parquet upload and Agent trace dataset workflows.
- Clearly warns that traces may contain prompts, paths, tool outputs, secrets, or PII.
- The SKILL.md provides no test results, test suite, or representative failure responses.
- It focuses on Dataset Viewer API workflows rather than being a complete data engineering or processing framework.
- Dataset creation and upload depend on additional CLI tools or browser actions, and error handling is not documented.
- Pricing, rate limits, and API availability are not addressed.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Hugging Face Dataset Explorer this page | 49 · Use with care | ★ 11k | 18d ago | Apache-2.0 |
| Trackio Experiment Tracking Skill | 53 · Use with care | ★ 38k | 14d ago | MIT |
| Hugging Face Datasets Skill | 50 · Use with care | ★ 38k | 14d ago | MIT |
| TAO AOI Image Mining ✓ NVIDIA · Official | 45 · Not recommended | ★ 3.6k | 12d ago | Apache-2.0 |
| cuPyNumeric Parallel Shard Loader ✓ NVIDIA · Official | 54 · Use with care | ★ 3.6k | 12d ago | Apache-2.0 |
The README presents hf-cli as the recommended entry point for broader Hugging Face Hub operations, including model search and dataset management. This skill has a narrower focus: Dataset Viewer API exploration, extraction, and related Parquet operations.
How did FollowSkills review this skill?
The evidence makes the core Dataset Viewer workflow read-only and advises using a Bearer token for gated/private datasets; defaulting trace uploads to private reduces exposure risk. However, the skill also includes dataset creation, uploads, and agent-trace uploads without per-operation confirmation, least-privilege token guidance, rollback, or pre-upload secret/PII scanning and redaction, so points are deducted.
The core endpoints, pagination parameters, and examples are reasonably consistent, and the document explains using total/page/partial fields for continuation. It lacks endpoint compatibility details, error-response handling, retry/rate-limit guidance, input validation, and skill-specific tests. Static evidence cannot exceed the 10-point cap, so points are deducted.
The skill clearly covers Dataset Viewer metadata, pagination, search, filtering, parquet, and statistics use cases. Trigger conditions, input/output boundaries, failure and non-fit cases are under-specified, and there is no evidence about Chinese-language support or mainland-China reachability. Upload guidance also mixes scope with the read-only workflow, so points are deducted.
The document has frontmatter, a staged workflow, defaults, endpoint examples, and some security guidance; repository evidence includes an Apache-2.0 license, CI generation checks, and a security reporting channel. The skill itself lacks versioning, changelog, maintenance ownership, update path, FAQs, and systematic limitation notes, while its external hf-cli dependency is not clearly documented as an installation prerequisite.
For common Dataset Viewer reading, pagination, search, and statistics tasks, the skill provides directly reusable curl templates and a clear core benefit. Representative results were not verified, API behavior and field assumptions still require user review, and the upload flow depends on other tooling. Static calibration caps this at 7, so full marks are not justified.
Evidence consists of the revision-pinned skill text plus repository CI, security-policy, and license material. The CI checks generated artifacts and version bumps but does not cover the skill's key API paths or result correctness. There is no third-party execution evidence or dedicated test suite, so verifiability is limited.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →