Hugging Face datasets skill: explore Hub datasets from your agent

Source: huggingface/skills · Maintainer: community: Hugging Face · License: Apache-2.0 · Checked: 24 October 2026

What is the Hugging Face datasets skill?

huggingface-datasets is Hugging Face's Agent Skill for exploring datasets on the Hub through the Dataset Viewer API. It lets Claude Code, Codex or another agent check a dataset's structure, read and search rows, filter them, and pull parquet links and statistics, without downloading the whole dataset first.

The skill's description reads: "Use this skill for Hugging Face Dataset Viewer API workflows that fetch subset/split metadata, paginate rows, search text, apply filters, download parquet URLs, and read size or statistics."

It comes from huggingface/skills, where the hf-cli skill is the recommended starting point. On skills.sh, huggingface-datasets had about 2,400 installs (snapshot, 24 October 2026), second in the repo after hf-cli.

What does it do?

The folder holds one SKILL.md with three parts: read-only exploration, creating and uploading datasets, and uploading agent traces.

Exploration. The core workflow tells the agent to:

1. Optionally check the dataset with /is-valid.
2. Resolve the subset (config) and split with /splits.
3. Preview with /first-rows.
4. Page through rows with /rows, using offset and length.
5. Use /search for text and /filter for row predicates.
6. Get parquet links from /parquet and totals from /size and /statistics.

It also lists /croissant for Croissant metadata. Every call is a GET against https://datasets-server.huggingface.co, with URL-encoded parameters and a 0-based offset. Its pagination example reads the IMDB dataset 100 rows at a time:

curl "https://datasets-server.huggingface.co/rows?dataset=stanfordnlp/imdb&config=plain_text&split=train&offset=0&length=100"

When results are partial, it tells the agent to use the num_rows_total, num_rows_per_page and partial fields to decide whether to continue. It also says to keep searches and filters "read-only and side-effect free", and for SQL on parquet it hands off to the hf-cli skill's hf datasets parquet and hf datasets sql commands.

Creating and uploading datasets. Two routes, from fewest dependencies up:

• The Hub web UI: create the repo at huggingface.co/new-dataset and upload parquet files on the "Files and versions" page.
• The JavaScript CLI: set HF_TOKEN, then run npx -y @huggingface/hub upload datasets/<namespace>/<repo> ./local/parquet-folder data, adding --private for a private repo.

After either, it says to call /parquet to confirm the shards appear in the viewer.

Agent traces. The last section covers uploading raw session files from Claude Code (~/.claude/projects), Codex (~/.codex/sessions) and Pi (~/.pi/agent/sessions) to a dataset so the Hub's trace viewer can browse them. It tells the agent to "Default to private dataset repos because traces can contain prompts, file paths, tool outputs, secrets, or PII."

What does the Dataset Viewer API allow?

The skill is a summary; Hugging Face's Dataset Viewer docs are the reference. The limits that matter for an agent:

The filter docs say where uses SQL-style syntax: column names in double quotes, string values in single quotes, operators such as =, <> and >=, combined with AND, OR and NOT. Sorting uses orderby, with DESC for descending order. The search docs say search runs on DuckDB full text search with BM25 ranking and a Porter stemmer, which assumes English text.

/rows, /search and /filter work only on datasets with parquet exports. On access, the quickstart says private and gated datasets need your token in the request headers, and "The dataset viewer supports private datasets owned by a PRO user or an Enterprise Hub organization."

How do you install it?

Skills CLI. The command shown on skills.sh uses the skills CLI:

npx skills add https://github.com/huggingface/skills --skill huggingface-datasets

Skills go into the current project by default, or across projects with -g. In Claude Code that means .claude/skills/huggingface-datasets/ or ~/.claude/skills/huggingface-datasets/, the standard Claude Code skill locations.

hf CLI. The huggingface/skills README says the plugin marketplaces expose only hf-cli, and that additional workflow skills install with:

hf skills add <skill-name>

So hf skills add huggingface-datasets. The README also notes the Cursor marketplace entry "is intentionally limited to hf-cli".

Codex and others. For Codex, the README says to copy or symlink skill folders into .agents/skills, in the repo or your home directory. A Gemini CLI extension is also listed. Codex and Cursor read the same SKILL.md format; see Codex skills and Cursor skills.

How do you use it?

Name a dataset and say what you want to learn. The agent works out the subset and split, then queries the API.

Example prompts (suggestions, not quotes from the repo):

• What subsets and splits does stanfordnlp/imdb have, and how many rows in each?
• Show me 20 rows from the train split of this dataset where the label is 1.
• Search the train split for rows mentioning "refund" and summarize what they have in common.
• Get column statistics for the test split and flag any column with lots of nulls.
• Upload ./out/parquet to a new private dataset called my-org/support-tickets, then check it shows up in the viewer.

Requirements: curl or any HTTP client, which agents already have. For uploads, a Hugging Face account and token, plus Node.js for the npx @huggingface/hub route. Uploads write to your account, so read the agent's plan before it runs them.

Tips and limits

From the files and pages linked, not from testing:

• Security audits. As of 24 October 2026, skills.sh showed Pass from Gen Agent Trust Hub and Socket and Warn from Snyk, all audited 15 September 2026. Snyk rated it MEDIUM with one issue, W011, "Third-party content exposure detected (indirect prompt injection risk)", because the agent reads user-submitted rows and text through /rows and /search. Treat dataset text as data, not instructions.
• 100 rows per call. Larger reads mean many requests. For full-dataset analysis, use the parquet URLs with DuckDB or the hf-cli skill's SQL command instead.
• Big datasets give partial search results. Search and filter only index the first 5GB, and the response says partial: true when that happens.
• Not every dataset has a viewer. Row, search and filter endpoints need a parquet export; /is-valid tells you which features a dataset supports.
• Traces can leak secrets. Hugging Face's agent traces docs warn that trace files can include prompts, command output, local paths, secrets, private code and personal data. Keep trace datasets private unless you have reviewed them.
• Two spellings, one flag. The skill's trace example uses --type dataset on hf upload, while the Hub docs show --repo-type dataset. The hf CLI reference lists --type and --repo-type as the same option.
• Old names on skills.sh. Listings such as hugging-face-datasets and hugging-face-dataset-viewer also appear with fewer installs; the current README uses huggingface-datasets.
• License. Apache-2.0, so you can copy and modify it with the license notice kept.

Save it in your Kipt library

In Kipt, open Skills and paste the huggingface-datasets folder link into the import box. For a repo with many skills, Kipt shows a list to pick from, so you can bring in hf-cli at the same time, since this skill hands SQL work to it. You can also drop the folder or a .zip. Your skills live in your own private GitHub repo and install into Claude Code or Codex with one command:

npx skills add https://github.com/<you>/<repo>.git --all -g

Kipt is not affiliated with Hugging Face.

Similar skills

• Hugging Face skill: hf-cli, the repo's main skill, which covers downloads, uploads, jobs and the SQL commands this skill points to.
• Find skills skill: search skills.sh from your agent for more data and ML skills.
• Data analysis prompts: prompts for the analysis step once the agent has the rows.
• Claude Code plugins: how plugin marketplaces like Hugging Face's work.