← All notes

Summary

The Hub has three kinds of repository: model, dataset, and Space. You create one in the browser or with Git. A dataset's README.md is its dataset card.

Repositories

The web UI needs no extra install. The terminal needs Git, and Git-Xet if large files such as images or model weights go through Git. To push, sign in. The easiest path in the guide is the hf CLI (brew install hf or pip install hf, then hf auth login).

Open huggingface.co/new, choose an owner (you or an organization), a name, public or private, and an optional license. The Hub then asks for a model card. That card is best practice; a test repo can skip it. The walkthrough itself creates a model repo and points elsewhere for datasets and Spaces.

On the Files tab, Add file either writes a new file or uploads one, with a commit message. Either action can open a pull request instead of committing to the main branch. hf upload sends a file or a directory from the terminal. A dataset repository needs --repo-type dataset. The CLI handles large files with no extra setup.

Clone over HTTPS or SSH. Dataset clones use the /datasets/ path. A push, or a private repo, needs your SSH public key in settings if you use SSH. Track files larger than 10MB with git-xet, then add, commit, and push. History shows each commit's diff.

Dataset cards

The Hub renders README.md on the dataset page. Say what the data contains, how to use it, and any biases. Metadata is YAML at the top, between lines of three dashes: language codes, a display name, tags, a license identifier, and task categories. Tags filter the datasets listing, and a license from the documented keyword list is shown on the page. The Metadata UI fills these fields in the browser.

A link to a paper page, or to an arXiv abstract or PDF, becomes an arxiv: tag that opens the paper and can filter models citing it. The Hub guesses modality from the files. Force one with a tag such as audio, image, text, or video. A library tag such as datasets or pandas makes that loader appear on the page.

An RL environment dataset is still a normal dataset repo. Tag it rl-environment plus a framework (harbor, verifiers, openenv, nemo-gym, or agentenv). The tags add loading commands under "Use this dataset". They do not make the Hub run the environment.

Claims

• id: C1 claim: The Hub has three repository kinds, and this guide creates a model repository as the example. source: S1 evidence: "There are three kinds of repositories on the Hub"
• id: C2 claim: A new repository is created at huggingface.co/new. source: S1 evidence: "To create a new repository, visit huggingface.co/new"
• id: C3 claim: The hf CLI handles large uploads without extra setup. source: S1 evidence: "The hf CLI handles large files automatically"
• id: C4 claim: Files larger than 10MB should be tracked with git-xet. source: S1 evidence: "Those files should be tracked with git-xet"
• id: C5 claim: The dataset card is the repository README.md, shown on the dataset page. source: S2 evidence: "This file is called a dataset card"
• id: C6 claim: Discovery tags are YAML metadata at the top of README.md. source: S2 evidence: "Tags are defined in a YAML metadata section at the top of the README.md file."
• id: C7 claim: Framework tags do not make the Hub run an RL environment. source: S2 evidence: "cause the Hub to execute the environment."

@misc{iil_note_hykcsxbs,
  author = {{iinventedlight}},
  title = {{Hub repositories and dataset cards}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-03},
  url = {https://iinventedlight.com/library/notes/hf-hub-repos-dataset-cards/},
  note = {Permanent ID iil:note:hykcsxbs. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt. This page in Markdown: /library/notes/hf-hub-repos-dataset-cards.md.