Hugging Face Scraper — AI Models, Datasets & Spaces

---

📑 Contents

• At a glance
• What this Hugging Face scraper does
• Key features
• Who it is for
• Use cases
• How it works
• Step-by-step usage
• Input
• Output
• Proven results (local verification)
• Integrations
• Tips for better results
• Limitations
• FAQ
• Keywords
• Disclaimer

---

⚡ At a glance

What you collect	AI models, datasets and Spaces (demos & apps) from Hugging Face
Search by	Keyword(s), author or organization — or leave empty for a popular sample
Sort by	Most downloads · Most likes · Trending now · Newest first · Recently updated
Key fields	Downloads, likes, trending score, task, library, license, languages, tags, author, dates
Empty run	Still returns about 100 popular models — useful results with zero setup
Output formats	JSON, CSV, Excel, HTML and XML via the Apify Console export
Code required	None — fill in a short form and press Start
Typical run	100 records in seconds; larger targeted runs scale with Max results

---

🤗 What this Hugging Face scraper does

Hugging Face is where the open AI ecosystem publishes its work: hundreds of thousands of ML models, training datasets and interactive Spaces. Browsing that catalogue by hand is slow, and copying names, download counts and licenses into a spreadsheet does not scale.

This Hugging Face scraper turns the catalogue into a dataset you can actually work with. Tell it what you are looking for — for example all models matching “llama”, everything published by one organization, or the most liked Spaces right now — and it returns one clean record per item with the numbers and metadata that matter:

• how popular it is (downloads, likes, trending score),
• what it does (task / pipeline, library, SDK),
• whether you may use it (license, gated status),
• who published it (author), when it appeared and when it last changed,
• and a direct link back to its Hugging Face page.

Whether you are building an AI models dataset for analysis, tracking an LLM leaderboard, shortlisting open-source models for a product, or mapping which datasets power modern training runs, this actor gives you the raw material in minutes instead of days of manual browsing.

---

✨ Key features

• Three catalogues in one actor — switch between Models, Datasets and Spaces with a single dropdown. No separate tools to learn.
• Keyword search that merges cleanly — add several search terms and get one combined, de-duplicated result set.
• Author / organization filter — watch a single publisher such as a lab, a company or a prolific community member.
• Five sort orders — most downloads, most likes, trending now, newest first or recently updated, so the items you care about surface first.
• License extracted for you — the license is pulled out of the raw tags into its own field, ready for filtering and compliance checks.
• Languages kept as data — language information stays as a proper list, not a flattened string, so you can filter multilingual coverage easily.
• Works with an empty form — press Start without typing anything and still get about 100 popular models to explore.
• Sensible size control — cap a run at exactly the number of records you need, or run a targeted search without a cap.
• Clean, analysis-ready records — consistent field names across models, datasets and Spaces, with empty values left out instead of cluttering your export.
• Export anywhere — download JSON, CSV or Excel, or send results to your stack with Apify integrations and the API.

---

👥 Who it is for

• ML engineers and researchers shortlisting models or datasets for a project, a benchmark or a paper.
• AI product teams tracking competitors, trending techniques and new releases in their space.
• Data analysts building AI market datasets, trend reports and internal dashboards.
• Developer-relations and community teams following authors and organizations that matter to their ecosystem.
• Students and curious builders who want a real, explorable sample of the open AI catalogue without writing code.

---

💡 Use cases

1. Build an LLM shortlist for a new product

Search for your task (for example “text-generation”), sort by downloads, and compare candidates side by side: downloads and likes show adoption, the task and library fields show fit, and the license field shows what you are allowed to ship. Export to a spreadsheet and score the shortlist with your team instead of juggling dozens of browser tabs.

2. Track what is trending in open AI

Sort any catalogue by Trending now and run the actor on a schedule. Diff this week's export against last week's and you have a simple, repeatable pulse on which models, datasets and demos the community is rallying around — useful input for roadmaps, content and research planning.

3. Follow one lab or organization

Put an organization in the Author field (for example a model lab) and collect everything it has published. Sort by newest first to catch fresh releases, or by downloads to see which of its releases actually landed.

4. Assemble training-data candidates

Switch to Datasets, search by topic or task, and use downloads, likes, languages and license to triage candidates before you inspect them in detail. The description field helps you discard obvious mismatches early.

5. Discover demos worth trying (or learning from)

Spaces are working demos, not just weights. Sort Spaces by likes or trending score to find interfaces people love, see which SDK they use, and open the linked page to try them yourself.

6. Feed an internal AI catalogue or dashboard

Run the actor via the Apify API or on a schedule and pipe the dataset into your warehouse, BI tool or internal search. Consistent fields across runs make week-over-week comparisons straightforward.

7. Support academic and market research

Build a citable snapshot of the open model ecosystem at a point in time: counts by task, library, license and language, adoption distributions, and release velocity — all from one export instead of manual sampling.

---

⚙️ How it works

1. You describe what you want — pick Models, Datasets or Spaces, optionally add keywords and an author, and choose a sort order.
2. The actor collects matching listings from Hugging Face's public catalogue, page by page, until it reaches your result cap.
3. Records are tidied as they arrive — duplicates are removed, the license and languages are lifted into their own fields, and each item gets a direct link to its Hugging Face page.
4. You get a dataset in the Apify Console: preview it in the table, filter it, and export it as JSON, CSV or Excel — or pull it through the API.

No setup, no scripts and nothing to maintain on your side. If you leave every field empty, the actor simply collects a popular sample of about 100 models so you can see the output shape right away.

---

🚀 Step-by-step usage

1. Open the actor in the Apify Console and go to the Input tab.
2. Choose What to scrape: Models, Datasets or Spaces.
3. (Optional) Add one or more Search keywords, e.g. llama.
4. (Optional) Add an Author or organization to limit results to one publisher.
5. Pick Sort results by — “Most downloads” is a solid default.
6. Set Max results to the size you want (100 is a good first run).
7. Press Start and watch the live log and status counter.
8. When the run finishes, open the Dataset tab to preview records.
9. Use Export to download JSON, CSV or Excel — or open the Integrations tab to send data onward automatically.
10. Re-run any time; try scheduling weekly runs to build a time series.

---

📥 Input

Every field is optional. Start with the defaults, then narrow down.

Field	What it does	Example
What to scrape	Pick the catalogue: Models, Datasets or Spaces	models
Search keywords	One or more keywords; each is searched and merged without duplicates	llama, sentiment
Author or organization	Only items from one publisher	meta-llama
Sort results by	Most downloads, most likes, trending, newest or recently updated	downloads
Max results	Cap on records collected; 0 means no cap on targeted searches	100
Proxy configuration	Connection settings; the default is right for most runs	default

Example 1 — Instant popular sample (zero setup)

{}

Returns about 100 popular models, sorted by downloads.

Example 2 — LLM discovery by keyword

{
  "entityType": "models",
  "searchTerms": ["llama"],
  "sort": "downloads",
  "maxItems": 100
}

Example 3 — Watch one organization, trending first

{
  "entityType": "models",
  "author": "meta-llama",
  "sort": "trendingScore",
  "maxItems": 50
}

Example 4 — Dataset hunting

{
  "entityType": "datasets",
  "searchTerms": ["instruction tuning"],
  "sort": "downloads",
  "maxItems": 100
}

Example 5 — Demo discovery in Spaces

{
  "entityType": "spaces",
  "searchTerms": ["image generation"],
  "sort": "likes",
  "maxItems": 100
}

---

📤 Output

Each run produces one record per model, dataset or Space. Field names are the same across all three catalogues, so exports are easy to combine and compare. Empty values are simply left out of a record.

Coverage legend (from the local verification runs below): 🟢 present for nearly every record · 🟡 present when the publisher provides it · ⚪ present for a smaller share of records.

Identity & popularity

Field	Meaning	Coverage
entityType	models, datasets or spaces	🟢 30/30 · 20/20 · 20/20 in tests
id	Full name, e.g. meta-llama/Llama-3.1-8B-Instruct	🟢 every record
author	Publisher (user or organization)	🟢 every record
downloads	All-time downloads (models & datasets)	🟢 30/30 models · 20/20 datasets
likes	Community likes	🟢 every record
trendingScore	Trending score when ranked by trending	🟡 10/10 in a trending run; not shown for download ranking

Classification & licensing

Field	Meaning	Coverage
pipelineTag	Task, e.g. text-generation (models)	🟢 29/30 in model test
libraryName	Library, e.g. transformers (models)	🟢 27/30 in model test
sdk	App SDK, e.g. gradio, docker (Spaces)	🟢 20/20 in Spaces test
license	License lifted out of the tags, e.g. mit, apache-2.0	🟡 26/30 models · 12/20 datasets · 14/20 Spaces
languages	Language codes as a list, when declared	⚪ 7/20 datasets; rarely declared elsewhere
tags	Full tag list as published	🟢 every record

Content & provenance

Field	Meaning	Coverage
description	Publisher description (datasets & Spaces, shortened if very long)	🟡 13/20 datasets · 12/20 Spaces
gated	Access gating: open, manual approval or similar (models & datasets)	🟢 30/30 models · 20/20 datasets
sha	Current revision identifier	🟢 every record
createdAt	When the item was published	🟢 every record
lastModified	When the item last changed	🟢 every record
url	Direct link to the Hugging Face page	🟢 every record
scrapedAt	When this actor collected the record	🟢 every record

Example output record

{
  "entityType": "models",
  "id": "meta-llama/Llama-3.1-8B-Instruct",
  "author": "meta-llama",
  "downloads": 4864172,
  "likes": 8573,
  "pipelineTag": "text-generation",
  "libraryName": "transformers",
  "license": "llama3.1",
  "tags": ["transformers", "safetensors", "llama", "text-generation", "license:llama3.1"],
  "gated": "manual",
  "sha": "0e9e39f249a16976918f6564b8830bc894c89659",
  "createdAt": "2024-07-18T08:56:00.000Z",
  "lastModified": "2024-09-25T17:00:57.000Z",
  "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct",
  "scrapedAt": "2026-10-23T13:40:27.263Z"
}

---

✅ Proven results (local verification)

This actor was run locally before release. All numbers below are from those real runs — same code, same output shape you will receive.

Check	Input	Result
Keyword search	Models · llama · most downloads · max 30	30 records, 30 unique IDs
Author watch	Models · author meta-llama · trending · max 10	10 records, every author meta-llama, trending score on 10/10
Datasets sample	Datasets · most downloads · max 20	20 records with descriptions on 13/20
Spaces sample	Spaces · most likes · max 20	20 records, SDK on 20/20
Empty run	{} (nothing typed at all)	100 popular models
Pagination capacity	Model listing, page after page	1,100 models in 11 pages, more pages still available afterwards

Field fill highlights from those runs:

• Identity and links were complete: id, author, url, entityType and scrapedAt appeared on every record checked.
• Popularity was complete: downloads and likes appeared on every model and dataset record; likes on every Space.
• Dates and revision (createdAt, lastModified, sha) appeared on every record checked.
• Model classification was near-complete in the keyword test: task on 29 of 30 records, library on 27 of 30.
• License — the field teams usually clean by hand — was already separated out on 26 of 30 models, 12 of 20 datasets and 14 of 20 Spaces. The rest simply do not declare a license.
• Sample links were opened and answered successfully for all three catalogues: a model page, a dataset page and a Space page.

What this means for you: the core columns you will filter and sort on arrive filled. Optional columns reflect what publishers themselves provide — nothing is padded with invented values.

---

🔌 Integrations

Because results live in a standard Apify dataset, they flow into the tools you already use:

• Spreadsheets — export CSV or Excel, or connect Google Sheets to refresh a workbook on a schedule.
• Databases & warehouses — pull the dataset through the Apify API into Postgres, BigQuery, Snowflake or your internal catalogue.
• Automation platforms — trigger Make, Zapier or n8n workflows when a scheduled run finishes (for example: “new model from a watched lab”).
• Chat alerts — post a Slack or email digest of trending items after each weekly run.
• BI dashboards — feed Looker Studio, Tableau or Power BI from the exported data for an always-current AI landscape view.
• Custom code — call the actor from the Apify API / SDKs in your own pipeline and read structured JSON back.

A typical pattern: schedule the actor weekly with your keywords, then let an automation diff and summarize what changed.

---

💡 Tips for better results

• Start broad, then narrow. Run the empty sample first to see the output shape, then add keywords.
• Combine keywords thoughtfully. Each keyword is searched on its own; overlapping terms are de-duplicated, so synonyms are safe.
• Use the author filter for watchlists. One run per important lab beats one giant mixed export you have to slice later.
• Match the sort to the question. Downloads answer “what is adopted?”, trending answers “what is hot now?”, newest answers “what just shipped?”.
• Filter by license early. If you plan to ship commercially, export and filter the license column before deeper evaluation.
• Keep runs comparable. Same input on a schedule gives you a clean time series; changing inputs mid-series breaks comparisons.
• Raise Max results gradually. 100 records is fast and cheap for exploration; scale up once you know the query is right.

---

⚠️ Limitations

• Public listings only. The actor collects catalogue metadata for publicly visible models, datasets and Spaces. It does not download model weights, dataset files or Space source code.
• Gated items show metadata, not contents. For access-controlled items you still get counts, license and dates — the underlying files remain behind Hugging Face's own access flow.
• Coverage mirrors the publishers. If a publisher declares no license, description or languages, those fields stay empty. The actor never invents values to fill a column.
• Trending scores are contextual. A trending score is published alongside trending-ranked results; download- or date-ranked exports focus on those rankings instead.
• Very long descriptions are shortened to keep records readable; follow the url for the full publisher text.
• Counts move. Downloads, likes and trending scores change constantly — treat each run as a timestamped snapshot (scrapedAt).

---

❓ FAQ

Do I need a Hugging Face account or API key?

No. The actor works with public catalogue information. Just press Start.

What happens if I run it with no input at all?

You still get data: about 100 popular models sorted by downloads. It is the fastest way to see exactly what the output looks like.

Can I scrape datasets or Spaces instead of models?

Yes — change What to scrape to Datasets or Spaces. Keywords, author, sorting and the result cap work the same way, and the output fields keep the same names.

How do I get everything from one company or lab?

Put its name in Author or organization (for example meta-llama), leave keywords empty, and raise Max results — or set it to 0 for a targeted run without a cap.

What does 0 in Max results mean?

No cap for that run. On a targeted search the actor keeps going until the catalogue results run out. An empty, filter-less run is still capped at a fast sample size so a stray click never turns into a huge job.

Is the license field reliable enough for compliance work?

It reflects the license the publisher declares, lifted into its own field so you can filter and review it at scale. Treat it as your triage list and confirm the full license text on the item's page before making legal decisions.

Why is trendingScore missing in my export?

Trending scores travel with trending-ranked listings. Sort by Trending now when that number matters; other rankings omit it by design rather than filling it with a stale value.

Can I run this on a schedule and track changes?

Yes. Schedule identical inputs (weekly works well) and compare exports by id and scrapedAt. New IDs are new releases; changing counts show momentum.

Which export format should I choose?

CSV or Excel for spreadsheet work, JSON for code and pipelines. The Console preview table is handy for a quick look and light filtering.

Does it download the actual models or datasets?

No — it collects the catalogue record (metadata, counts, license, links). That keeps runs fast and storage small. Use each record's url to open the item and obtain files through Hugging Face itself.

---

🔍 Keywords

Hugging Face scraper · Hugging Face models scraper · AI models dataset · ML model discovery · LLM leaderboard data · open-source LLM list · Hugging Face datasets export · Hugging Face Spaces scraper · transformers model catalogue · model license checker data · AI market research data · trending AI models · LLM comparison data · machine learning dataset discovery

---

📄 Disclaimer

This is an independent tool for collecting publicly available catalogue information. Not affiliated with Hugging Face. Hugging Face is a trademark of its respective owner; product and organization names appear here only to describe what the data refers to. Please respect Hugging Face's terms and each publisher's license when using collected data.