Explore Hugging Face Datasets via API

Provides read-only exploration of a Hugging Face dataset through the Dataset Viewer API (base URL datasets-server.huggingface.co): validate availability (/is-valid), resolve config and split (/splits), preview (/first-rows), paginate rows (/rows with offset/length, max 100), full-text search (/search) and predicate filtering (/filter), and retrieve parquet shard links (/parquet) plus size and column statistics (/size, /statistics) and Croissant metadata (/croissant) when available. Gated or private datasets require an Authorization: Bearer <HF_TOKEN> header.

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

For SQL-style querying, it derives Hub parquet alias paths (hf://datasets/<namespace>/<repo>@~parquet/<config>/<split>/<shard>.parquet) from the /parquet endpoint and runs them through npx parquetlens, including COPY ... TO exports to CSV, JSON, or Parquet. For creating and uploading datasets, it offers a zero-dependency flow - create the repo and upload parquet files via the Hub UI, then verify shards via /parquet - or a low-dependency CLI flow (npx @huggingface/hub upload, with a --private flag for a private repo on creation), reading the auth token from HF_TOKEN.

Use it for read-only exploration and extraction of a Hugging Face dataset - validating, previewing, paginating, searching, filtering, or SQL-querying its rows - keeping all Dataset Viewer calls side-effect free.

Input: a dataset namespace/repo identifier, plus config, split, and pagination or filter parameters. Output: JSON rows, search or filter matches, parquet shard links, size/statistics metadata, or a SQL query result exported to CSV, JSON, or Parquet.

Runs against the Dataset Viewer API (datasets-server.huggingface.co) and the Hugging Face Hub (huggingface.co/new-dataset, the @huggingface/hub CLI), with parquetlens for local SQL querying of Hub-hosted parquet files.

Developers exploring, querying, or uploading Hugging Face datasets who want direct API and CLI workflows instead of downloading full datasets locally.

Its core workflow is sequential: optionally validate the dataset with /is-valid, resolve the config and split with /splits, preview with /first-rows, then paginate with /rows before reaching for /search or /filter. When pagination is partial, response fields like num_rows_total, num_rows_per_page, and partial drive continuation logic, and /search matches string columns while /filter requires predicate syntax in a where parameter with optional sorting via orderby - both calls stay strictly read-only and side-effect free. Query parameters must be URL-encoded, offset is 0-based, and the API's default HTTP method throughout is a plain GET request, with a length cap of 100 rows per page on row-like endpoints.

Uploading a parquet folder via the CLI auto-creates the destination dataset repo if it doesn't already exist, and after any upload, calling /parquet again discovers the resulting config/split/shard values needed for querying that same data back through the @~parquet alias.