Hugging Face Datasets
The Hugging Face Hub, run by the company Hugging Face, hosts datasets used to train and evaluate machine-learning models. Through ScholarMuse, an agent searches the Hub's dataset listing by keyword and gets each dataset's repo id, owner, licence tags, task categories, modality, format, size category, download and like counts, and whether access is gated.
ScholarMuse tools: search_datasets. Each successful call uses credits.
The web tool runs without an account, up to 5 runs a day per visitor.
Ask your agent
What your agent gets back
Every record carries id, source, title, url, license, attribution and retrievedAt. The fields below are the data that comes from Hugging Face Datasets.
search_datasets
The record title is the repo id, such as owner/name. license, size, format, modality and tasks are read from the dataset's Hub tags and are missing when the owner did not set them.
- creators
- Owner account of the dataset repo
- year
- Year the repo was created
- lastModified
- Date of last change (YYYY-MM-DD)
- description
- Dataset description, cut at 2,000 characters
- size
- Size category tag set on the Hub
- format
- Format tags, such as parquet
- modality
- Modality tags, such as text or audio
- tasks
- Task category tags, up to 6
- downloads
- Download count reported by the Hub
- likes
- Number of likes on the Hub
- gated
- True when access must be requested
- license
- Licence tags of the dataset itself
What it covers
- Dataset repositories listed on the Hugging Face Hub, mostly machine-learning training and evaluation data.
- Results are ordered by download count, highest first, not by relevance.
- Each record links to the dataset card on huggingface.co.
Limits
- One call returns at most 25 datasets (default 10). When several repositories are searched together, that number is shared between them, so name Hugging Face alone to fill it.
- The only search input is a keyword query. There are no year, licence, author or file-type filters; the agent has to read those fields in the results.
- Because results are ranked by downloads, a little-used dataset can fall outside the 25 returned.
- Rows, column schemas and files are not returned, and model or Space repos are not searched. An upstream request is abandoned after 30 seconds.
How ScholarMuse reaches Hugging Face Datasets
The connector calls huggingface.co/api/datasets, sorted by downloads. It sends a Hugging Face token when one is configured on our server and works without one.
Responses are cached for up to 6 hours, so a dataset published at Hugging Face in the last few hours may not appear yet.
API key: the upstream API can be called without one, so you do not add one.
Licence and attribution
Licence: Hugging Face Terms of Service (listing metadata)
Listing from the Hugging Face Hub (huggingface.co). Each dataset carries its own licence; check the dataset card.
Records that come from Hugging Face Datasets carry this licence and attribution. When you cite the data in a paper, follow Hugging Face Datasets's own citation guidance.
