Hugging Face models + datasets - search, metadata (Hub API)

Returns: one row per Hugging Face model or dataset: id, author, name, task (pipeline tag / task category), library, license, languages, downloads (30-day), likes, trending score, last modified, created, gated/private flags, file count, tags, description (datasets), Hub URL. Input: { "kind": "models", "search": "llama", "task": "text-generation", "license": "apache-2.0", "sort": "downloads", "max_records": 100 } — all optional; empty {} returns the 100 most-downloaded models. Price: $5.00 per 1,000 rows ($0.005 per row) plus $0.10 per run. Pay per event; no subscription.

Search the Hugging Face Hub — models or datasets — by term, task, library, license, language, author or any Hub tag, sorted by downloads, likes, recency or trending score, and get a flat, normalised row per repo, ready for CSV or JSON: model landscape tables, license audits, "what is popular for task X", dataset inventories. No token needed: it reads the public huggingface.co/api endpoints the Hub website uses. An optional read token raises the Hub's rate limits.

The default input (models matching llama, most downloaded first, 100 rows) succeeds within 5 minutes and returns rows — Apify auto-tests it daily.

Data source and status

Publisher	Hugging Face, Inc. — the Hub (https://huggingface.co)
API	GET /api/models and GET /api/datasets with search, author, filter (comma-separated ANDed tags), sort, direction=-1, limit=100, full=true. Public, no key. Documentation: https://huggingface.co/docs/hub/api
What this Actor reads	Plain GET requests, at most one per request_interval_ms (default 500 ms); paging follows the API's Link: rel="next" cursor header
Refresh cadence	Live on every run — downloads is the Hub's rolling 30-day figure, likes and last_modified are current; fetched_at on each row is the fetch time
Terms	Hugging Face Terms of Service and Hub API guidelines apply (https://huggingface.co/terms-of-service). Repo metadata is public; model and dataset contents carry their own licenses (the license column) — this Actor never downloads files
Known caveats	The API gives no total match count; task is empty for models without a pipeline tag; license comes from cardData.license or the license: tag and is empty when the author set none; datasets have no library (the format: tag is used instead) and carry a description (first 500 chars of the card); gated is false, true, "auto" or "manual"; search is a substring match on the repo id, not full-text

Honesty note: the rows are the Hub API's own fields renamed and flattened — nothing is inferred, summarised or classified by this Actor, and a run stops with an error rather than guess when the API's shape changes.

Identification: the requests carry a product-token User-Agent and the operator's contact address in the standard From: header (RFC 9110 §10.1.2).

Input

Field	Type	Meaning
kind	models / datasets	Which catalogue (default models)
search	string	Substring on the repo id, e.g. llama, whisper, squad
task	string	Models: pipeline tag (text-generation, text-classification, ...). Datasets: task category (question-answering, ...)
library	string	transformers, pytorch, safetensors, diffusers, gguf, ...
license	string	apache-2.0, mit, llama3, cc-by-4.0, ...
language	string	ISO code, e.g. en
author	string	Hub user or organisation, e.g. meta-llama
extra_filters	array	Further Hub tags, ANDed
sort	string	downloads (default), likes, lastModified, createdAt, trendingScore
max_records	integer	Default 100, at most 20,000
max_tags	integer	Tags kept per row (default 40; tag_count carries the total)
hf_token	secret string	Optional hf_... read token — raises rate limits, never logged
request_interval_ms	integer	Default 500 (floor 200)
contact_email	string	Sent in the From: header

Example — the 30 most-liked English question-answering datasets under the MIT license:

{ "kind": "datasets", "task": "question-answering", "license": "mit", "language": "en", "sort": "likes", "max_records": 30 }

Output (dataset row)

id, kind (model / dataset), author, name, task, library, license, languages, downloads, likes, trending_score, last_modified, created_at, gated, private, disabled, sha, file_count, tag_count, tags, description, url, api_url, fetched_at, source.

A run summary (filters, pages, requests, whether a token was used, whether the pay-per-event budget stopped the run) is stored as RUN_SUMMARY in the run's key-value store.

Pricing (pay per event)

Event	Price
run-start — once per run	$0.10
record — per row written	$0.005

The default run (100 rows) costs $0.60; 1,000 rows cost $5.10. Rows stop when your run's maximum charge is reached; the run summary says so.

Operator

Steelyard Ventures LLC (RetrainMap) — info@steelyardclinical.com. Not affiliated with Hugging Face, Inc. The Actor writes only to its own dataset and key-value store; the optional token is used for the Hub requests only and is never stored or logged.