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.
This skill supports Hugging Face Dataset Viewer API workflows for exploring and extracting dataset content without needing to download entire datasets. It solves the problem of programmatically inspecting large datasets on the Hub, covering how to fetch subset and split metadata, paginate rows, search text, apply row filters, discover parquet shard URLs, and read size or column statistics through documented read-only GET endpoints.
The core workflow validates a dataset with /is-valid, resolves config and split via /splits, previews with /first-rows, paginates content through /rows using offset and length (max 100), then uses /search for text matching and /filter for predicates, and retrieves parquet links plus totals via /parquet, /size, and /statistics. It documents defaults such as the base URL, GET method, URL-encoding, 0-based offset, and that gated or private datasets require an Authorization bearer token. Beyond read access it also describes creating and uploading datasets through the Hub UI or a low-dependency CLI flow using npx @huggingface/hub, setting an HF_TOKEN environment variable, and uploading parquet folders, optionally as private repos. A section on agent traces explains uploading raw JSONL session traces from Claude Code, Codex, or Pi Agent, and sensibly defaults to private dataset repos because traces can contain prompts, secrets, or PII.
It targets data scientists, ML engineers, and dataset maintainers who work with Hugging Face datasets. Typical use cases include previewing and paginating dataset rows, searching and filtering content, discovering parquet URLs for downstream querying, and publishing new parquet-based datasets to the Hub. The emphasis throughout is on read-only, side-effect-free exploration, with credential handling via a standard environment variable.
The Dataset Viewer endpoints (rows, search, filter, size, statistics, parquet) are read-only and side-effect free. The skill separately documents optional upload flows for creating and publishing datasets, which do write to the Hub.
Gated or private datasets require an Authorization bearer token, and the upload flows use an HF_TOKEN environment variable exported in the shell. These are standard Hugging Face credential conventions; no exfiltration is involved.
Row-like endpoints use offset (0-based) and length, with length capped at 100. Response fields such as num_rows_total, num_rows_per_page, and partial drive continuation logic for large datasets.
It supports uploading raw JSONL traces from Claude Code, Codex, or Pi Agent, and defaults to private dataset repos because traces can contain prompts, file paths, tool outputs, secrets, or PII. It advises preserving raw files and nesting by project rather than dumping everything at the repo root.
Quick Setup:
.claude/skills/Repository
huggingface/skills