Datasets on Hugging Face

The Datasets section shows the datasets of your Hugging Face account. The app keeps no copy of them: every page is read from the Hub with your own token, at one precise commit. Generated data that is not published yet is listed under Drafts instead (see Getting Started).

The list

The list shows the dataset repos of the namespace saved in Settings, or of your token's account when none is saved, private repos included. When your token belongs to organizations, a menu switches to one of them.

Search matches the name, description and tags. The list filters on visibility (all, private, public) and sorts by last update, name, downloads, likes or size.

A dataset's page

Tab	What it shows
Data	The rows of a split, 25, 50 or 100 per page. Search covers every text field, nested ones included. Filters, sorting by column, Q/A datasets shown as cards. A click opens the whole row.
Card	The README rendered, then its metadata. Raw HTML is not rendered and images become links.
Statistics	Per column: type, empty cells, value counts for label-like columns, histograms of text lengths, list sizes and numbers. Click a value to see its rows. Q/A datasets also count repeated questions.
Files	Every file with its size and the split it feeds, with links to the Hub.
History	Every commit is a version. Browse shows the data as of that commit; Changes compares it with the previous one, row by row.

In the row panel, ← and → move to the previous or next row and Esc closes it. Filters accept is, is not, contains, <, ≤, >, ≥, is empty and is not empty, depending on the column's type.

The view lives in the URL (tab, version, split, search, filters, page, open row): copy the address to share exactly what you see.

Build a collection embeds a split of the dataset into Qdrant, to search it and test its retrieval: see Vector collections.

Q/A datasets

A split is shown as Q/A cards when it has a question column (question, query, prompt, instruction or q) and an answer column (answer, response, output, completion or a). A context, source_text or passage column is shown under the answer.

Comparing versions

Changes matches rows by their id column when every row of both versions has a unique one: an edited answer then counts as changed. Without such a column, rows are matched by content, so an edit counts as one removal plus one addition. The first 20 rows of each kind are listed.

How splits are found

1. The configs block of the card's YAML header, when it has one.
2. Otherwise the Hub's naming rules: a file or folder named train/training, validation/valid/val/dev or test/testing/eval/evaluation feeds that split.
3. Otherwise every data file of the most common format forms the train split.

The viewer reads JSONL, JSON, CSV, TSV and Parquet. A JSONL, CSV, TSV or Parquet file outside every split, such as the review.jsonl the GitHub job holds out for review, is offered as an extra split.

Limits

Limit	Value	Setting (apps/server/core/config.py)
Largest split the viewer loads	50 MB of data files (else: open the Hub)	max_hub_view_bytes
Loaded splits kept in memory	256 MB in total, per commit	hub_cache_bytes
Reuse of a revision → commit lookup	30 seconds, per token	hub_resolve_ttl_seconds
Text cell in a page of rows	5,000 characters (the row panel: all)	MAX_CELL_CHARS (services/hub.py)
Versions listed	The 100 most recent commits	MAX_COMMITS (services/hub.py)

Why the app reads the files itself

The Hub's own dataset viewer API serves private datasets only to PRO accounts and Enterprise organizations (dataset viewer docs), and every dataset this app publishes is private.

API

Every route reads with the caller's own Hugging Face token, so the Hub decides what each user sees: a repo the token can't read answers 404, like an unknown one.

Route	Returns
GET /hub/datasets	The dataset repos of one of your namespaces (author)
GET /hub/datasets/{owner}/{name}	Metadata, card, files, splits and refs at one commit
GET /hub/datasets/{owner}/{name}/rows	A page of rows, with search, filters and sort
GET /hub/datasets/{owner}/{name}/row	One row in full (row_idx)
GET /hub/datasets/{owner}/{name}/statistics	The per-column statistics of a split
GET /hub/datasets/{owner}/{name}/commits	The versions, newest first
GET /hub/datasets/{owner}/{name}/compare	Rows added, removed and changed between base and head

rows takes revision, config, split, offset, length (up to 100), q, filters, sort and order. filters is a JSON list:

[{ "column": "category", "op": "eq", "value": "docs" }, { "column": "confidence", "op": "gte", "value": 0.5 }]

Errors: 503 without a Hugging Face token, 404 for an unknown repo, revision, split or row, 413 for a split above the size limit, 400 for a filter or sort that doesn't apply, 422 for data files that don't read as a table, 502 when the Hub fails.