Hugging Face Dataset Metadata Scraper (Hub Downloads and Likes)

Extract public Hugging Face dataset metadata from the official Hub API: dataset id, author, last-modified date, downloads, likes, license, gated/disabled flags and structured tags as JSON. One dataset per result, schema-validated, with no dataset card text and no dataset files.

What it does

Query the official Hugging Face Hub datasets API (https://huggingface.co/api/datasets), the public JSON endpoint behind the Hub catalogue, and get one dataset item per dataset. Filter by keyword, author or library, sort by last modification, downloads or likes, and collect up to 5,000 datasets per run.

The source is a documented public API, not a scraped web page: plain HTTP, no browser, no login, no CAPTCHA and no proxy. The Actor reads only metadata; it never downloads dataset files and never reads the dataset card body.

Input: keyword, author, library, sort (lastModified, downloads or likes; default lastModified) and maxResults (default 100, maximum 5,000), plus mode and strictSchema.

Output: one dataset item with dataset_id, author, last_modified, downloads, likes, license, gated, disabled, tags and source_url, plus the common url, title and scrapedAt envelope. Format: JSON dataset.

mode=search (default) emits every match up to maxResults. mode=monitor emits only datasets that are new or whose last_modified moved forward since the previous monitor run, deduplicated by dataset_id; the state is kept in the run's key-value store.

Why the output carries no card text

A Hugging Face dataset card can contain a person's email, Telegram or X handle in its free-text description. The API response exposes description and cardData, but this Actor never maps them: only the structured metadata fields above can reach the dataset. Any tag that looks like a personal contact is dropped as well. See the FAQ and the automated GDPR tests.

Example output

Real sample read from the Hub API on 2026-10-20 and produced by the committed fixture (datasocial/tiktok-5.6B-videos):

{
    "url": "https://huggingface.co/datasets/datasocial/tiktok-5.6B-videos",
    "title": "datasocial/tiktok-5.6B-videos",
    "scrapedAt": "2026-10-20T12:00:00.000Z",
    "dataset_id": "datasocial/tiktok-5.6B-videos",
    "author": "datasocial",
    "last_modified": "2026-10-07",
    "downloads": 25678,
    "likes": 377,
    "license": "cc-by-nc-4.0",
    "gated": false,
    "disabled": false,
    "tags": [
        "license:cc-by-nc-4.0",
        "size_categories:1B<n<10B",
        "format:parquet",
        "modality:tabular",
        "modality:text",
        "library:datasets",
        "library:dask",
        "library:polars",
        "library:mlcroissant",
        "region:us",
        "tiktok",
        "social-media",
        "short-video",
        "creators"
    ],
    "source_url": "https://huggingface.co/datasets/datasocial/tiktok-5.6B-videos"
}

Pricing

Pay-per-event pricing (.actor/pay_per_event.json, referenced from actor.json as payPerEvent):

Event	Charged when	Price
apify-actor-start	a run starts (synthetic, automatic)	0.00005 USD per run (Apify-recommended default; includes 5 seconds of free compute)
apify-default-dataset-item	a validated dataset is saved to the dataset (synthetic, automatic)	0.003 USD per result = 3.00 USD per 1,000 results

There is no custom per-item event and no Actor.charge() call: the platform already charges apify-default-dataset-item per pushed item, so another event would bill users twice. Malformed records, empty searches and repeated datasets in monitor mode are never pushed, therefore never charged.

The price sits inside the direct-competitor band reported for this niche (0.60-32 USD per 1,000) and inside Steadfetch's 1-10 USD band.

Measured cost per 1,000 results (no browser, no proxy, public API): a 1,000-record page measured 0.96 MB and 0.2-0.4 s against the live API on 2026-10-20. At 256 MB and Apify's USD 0.20 per GB-hour compute rate that is about USD 0.00002 per 1,000 results; a conservative budget including run overhead and retries is USD 0.01 per 1,000 results. Gross margin at 3.00 USD per 1,000 is above 99%. The first paid platform run will replace this estimate with the billed figure.

FAQ

Do I need a Hugging Face account or API key? No. The Hub datasets API used here is public and unauthenticated. There is no token input and no secret to configure.

Does it need a browser? No. The Hub returns structured JSON; plain HTTP is roughly 10x cheaper and faster than a browser-based Actor.

Does it download the datasets? No. Only metadata is read. Dataset files and the dataset card body are never fetched.

Does it return personal data? No. description and cardData are never mapped, and tags that look like personal contacts are dropped. The automated tests feed a record with an email, a Telegram handle and an X handle and assert that none of them appears in the output.

What is monitor mode? mode=monitor keeps state between runs in the key-value store and emits only datasets that are new or have a newer last_modified date, deduplicated by dataset_id. It is most complete with the default sort=lastModified. Re-running it without changes emits and charges nothing.

How do I find datasets from one organisation? Set author to the Hugging Face user or organisation name, for example datasocial.

When am I charged? Once per run start (apify-actor-start) and once per valid dataset saved (apify-default-dataset-item). Errors, empty results and duplicates are not charged per result.

Is the output schema stable? Yes. Required fields in .actor/dataset_schema.json never change; new fields are added additively and the item schema allows extra properties.

Maintenance

Maintenance status: tested daily. The Actor is registered in the Steadfetch daily catalog QA workflow with the minimal input {"keyword":"medical","maxResults":10}; the label Tested daily. Last verified: {YYYY-MM-DD}. is published only after the first successful scheduled run is logged.

The source is the versioned public Hub API. Committed fixtures and schema tests catch response changes in CI, and any break in the daily run is reported as an issue in the Steadfetch Actors repository.

Keywords

hugging face datasets, dataset metadata, dataset search, dataset catalogue, machine learning datasets, downloads, likes, license, JSON dataset, ML data

Input reference

Search mode:

{
    "keyword": "medical",
    "maxResults": 100,
    "sort": "lastModified"
}

Monitor mode:

{
    "mode": "monitor",
    "author": "datasocial",
    "maxResults": 100,
    "sort": "lastModified"
}

Field	Type	Description
mode	search or monitor	Default search. monitor emits only new or updated datasets since the previous run.
keyword	string	Optional Hub search term.
author	string	Optional Hugging Face user or organisation.
library	string	Optional Hub library filter, for example datasets, pandas, polars or dask.
sort	lastModified, downloads or likes	Default lastModified; highest/newest first.
maxResults	integer	Default 100, maximum 5,000.
strictSchema	boolean	Default true; fail the run on schema violations.

Development

npm install
npm test                # lint + typecheck + offline Vitest suites (no credit)
npm run publish:check   # Steadfetch publish gate
npm run build && npm run verify:build   # production entrypoint guard
npm run schemas:validate
RUN_LIVE_TESTS=1 npm test               # also runs the live Hub API tests (no credit)

The default test suite runs against committed fixtures and never touches the external network or Apify. Live tests are skipped unless RUN_LIVE_TESTS=1; they validate real Hub responses against the same output schema. For a credit-free end-to-end run, point the Actor at a loopback server serving test/fixtures/datasets-response.json and run apify run --input ....

Project structure:

.actor/actor.json          Actor definition (payPerEvent, dataset schema, output schema)
.actor/input_schema.json   INPUT options (mode, keyword, author, library, sort, maxResults)
.actor/dataset_schema.json Item JSON Schema (fields) + table view for the Store
.actor/pay_per_event.json  PPE pricing (two synthetic platform events)
src/fetch.ts               Hub API client: filters, cursor pagination, retry/backoff
src/extract.ts             Pure metadata mapping, GDPR tag scrubbing and monitor dedupe
src/main.ts                Entry point: fetch, extract, monitor filter, validate, push
src/validate-output.ts     Ajv validator shared by the run and the tests
test/                      Offline fixtures, Vitest suites and optional live API tests