Hugging Face Scraper: Models, Datasets & Spaces

Export Hugging Face Hub data to JSON, CSV or Excel. Everything comes from the official Hub API — the same JSON the website itself calls — so there is no HTML parsing and no login for public repositories.

One Actor, four jobs:

Mode	What you get
Models	Browse or search models with downloads, likes, trending score, task, library, licence, base model, languages and dates
Datasets	The same for datasets, plus description, task categories, size category and file formats
Spaces	The same for Spaces, plus SDK, title, short description and the models a Space runs on
Specific repositories	The full record for a list of repositories you name, including parameter count, storage size, model type and the Spaces using a model

Why use this one

• The Hub hides half its metadata inside tags. Licence, language, base model, arXiv paper, training dataset, task category and size category all arrive as namespaced strings in one tags array. This Actor splits them into license, languages, baseModel, arxivIds, datasetsUsed, taskCategories and sizeCategory, so a spreadsheet can sort and filter on them.
• Parameter counts as a number. parameterCount comes from the repository's safetensors index, so you can filter for models that will actually fit on your GPU instead of guessing from the name.
• Base model lineage. baseModel tells you what a fine-tune was built on — the quickest way to find every derivative of a given model.
• Real pagination. The Hub pages its list endpoints with a cursor in the Link header. This Actor follows it, so a request for 2,000 models returns 2,000 rows rather than the first 100.
• No token needed. Public models, datasets and Spaces are readable anonymously.

How to use it

1. Choose What to scrape.
2. For browsing, set any of Search terms, Authors or organisations and Filter tag, then pick a sort order.
3. For Specific repositories, list them as owner/name and set Repository type.
4. Click Start, then download the results or pull them through the Apify API.

Example input: the most downloaded text-generation models

{
  "mode": "models",
  "filterTag": "text-generation",
  "sortBy": "downloads",
  "maxResults": 500
}

Example input: everything an organisation publishes

{
  "mode": "models",
  "authors": ["meta-llama", "mistralai", "Qwen"],
  "sortBy": "lastModified",
  "maxResults": 200
}

Example input: trending Spaces this week

{
  "mode": "spaces",
  "sortBy": "trendingScore",
  "maxResults": 100
}

Example input: a watchlist of specific models

{
  "mode": "repoDetails",
  "repoType": "models",
  "repositories": [
    "openai-community/gpt2",
    "meta-llama/Llama-3.1-8B-Instruct",
    "https://huggingface.co/BAAI/bge-m3"
  ]
}

Output

Each item carries a type field: model, dataset or space.

Model

{
  "type": "model",
  "hub": "huggingface",
  "repoId": "openai-community/gpt2",
  "name": "gpt2",
  "author": "openai-community",
  "pipelineTag": "text-generation",
  "libraryName": "transformers",
  "downloads": 15870518,
  "likes": 4248,
  "license": "mit",
  "languages": ["en"],
  "baseModel": null,
  "parameterCount": 137022720,
  "modelType": "gpt2",
  "fileCount": 32,
  "tags": ["transformers", "pytorch", "safetensors", "gpt2"],
  "createdAt": "2022-03-02T23:29:04Z",
  "lastModified": "2024-02-19T10:57:45Z",
  "url": "https://huggingface.co/openai-community/gpt2"
}

Space

{
  "type": "space",
  "hub": "huggingface",
  "repoId": "gradio/hello_world",
  "author": "gradio",
  "sdk": "gradio",
  "title": "Hello World",
  "likes": 112,
  "runtimeStage": "RUNNING",
  "hardware": "cpu-basic",
  "models": [],
  "appUrl": "https://gradio-hello-world.hf.space",
  "url": "https://huggingface.co/spaces/gradio/hello_world"
}

Pricing

Pay only for results:

Result	Price
Repository from a browse or search	$3.00 per 1,000
Repository from Specific repositories	$5.00 per 1,000

Set a maximum spend per run in the run options and the Actor stops cleanly when it is reached.

Limits to know about

These come from the Hub API, not from this Actor:

• trendingScore is only on list results. The single-repository endpoint does not return it, so rows from Specific repositories have trendingScore: null. Sort by Trending in a browse mode to get it.
• runtimeStage and hardware are only on the single-Space endpoint. Rows from Spaces browsing leave them null; use Specific repositories with Spaces for live status.
• downloads counts the last 30 days, which is what the Hub publishes. There is no all-time figure in the API.
• Spaces have no download count. Likes and trending score are the only popularity signals the Hub exposes for them.
• parameterCount needs safetensors. Repositories published only as GGUF, PyTorch .bin or ONNX leave it null — nothing is inferred from the model name.
• Gated and private repositories need a token. Without one they are skipped and the reason is logged. Browsing never lists private repositories at all.
• One tag per filter. The Hub's filter parameter takes a single tag; run the Actor twice to intersect two of them.

Optional Hugging Face token

Leave Hugging Face token empty and the Actor reads public data anonymously, which is what most runs need. Adding a free read token from huggingface.co/settings/tokens does two things: it raises your rate limit on big crawls, and it lets the Actor read gated repositories that your account has already been granted access to.

The token field is encrypted by Apify, is sent only to huggingface.co, and is never written to the dataset or the log. A read-only token is enough — never paste a write token.

FAQ

Do I need a Hugging Face account? No, not for public repositories.

Can this download model weights? No. It exports metadata about repositories, not their files.

How do I find every fine-tune of a model? Browse models with the filter tag base_model:<owner>/<name>, then sort by downloads.

Why do fewer rows come back than I asked for? The Hub has run out of matches for that filter — check the log, which prints the count per search.

Can I track releases daily? Yes. Sort by Recently updated or Newest, create a schedule in Apify Console, and connect the dataset to Slack, Google Sheets, Zapier or Make.

Feedback

Found a bug or need a field that is missing? Open an issue on the Actor's Issues tab.