Explore Hugging Face Datasets via API

Hugging Face Dataset Viewer supports dataset exploration and extraction through read-only API calls. Its workflow begins with an optional availability check, resolves the dataset configuration and split, and previews initial rows before retrieving further content. This sequence connects a dataset identifier to the actual subset of data you want to inspect rather than assuming that every repository exposes a single configuration or split.

The skill covers text search, predicate filtering, parquet discovery, size totals, column statistics, and Croissant metadata when available. It uses the Dataset Viewer service at datasets-server.huggingface.co and defaults to GET requests. Query parameters need URL encoding, row offsets start at zero, and row-like endpoints usually accept a maximum length of 100. Apache-2.0 is the stated license.

Use this skill when a task falls within Hugging Face Dataset Viewer or the associated dataset preparation workflow. It fits inspection before extraction: determine whether a dataset is available, identify its split, examine representative rows, then decide whether to page through rows or discover parquet files.

Searches and filters should remain read-only and free of side effects. For CLI-based parquet discovery or SQL queries, the workflow directs you to the hf-cli skill and its dataset parquet and SQL commands. Before changes, verify commands, API behavior, credentials, quotas, pricing, and deployment effects against current official documentation. Generated examples do not replace environment-specific testing or approval for destructive or costly actions.

A dataset repository identifier supplies the starting input. Configuration and split values narrow row requests; offset and length control pagination. Search accepts text, while filtering accepts a predicate and optional ordering. The viewer returns rows or metadata appropriate to the endpoint, including parquet shard links and column statistics.

Pagination continues by changing the offset while preserving the dataset, configuration, and split:

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

When results are partial, response fields such as num_rows_total, num_rows_per_page, and partial guide continuation. Text search matches string columns; predicate filtering uses where and optional orderby parameters.

Gated and private datasets require bearer authentication with HF_TOKEN. Dataset preparation can use the Hub browser interface to create a repository and upload parquet files through Files and versions, followed by verification that viewer shards appear. A low-dependency alternative uses npx @huggingface/hub or hfjs to upload a local parquet folder, creating the repository when needed and optionally making it private.

The Hub also accepts original JSONL session traces and can detect their format, tag them as Traces, and enable browsing of sessions, turns, tool calls, and responses. Keep trace repositories private by default because traces can contain prompts, paths, outputs, secrets, or personal information. Preserve raw files and organize them by project or working directory rather than placing every session at the repository root.