Skip to main content

Dataset registry

The registry is where training datasets are shared between accounts. You will find it under Registry → Datasets, next to Images and Models.

It answers two questions:

  • What can I train on that I did not have to build?
  • What have I shared, and can I take it back?

Your own working datasets — the ones you are still uploading, validating and training against — stay under Compute → Datasets. Publishing is the only thing that moves a dataset between the two.

Browsing

The Community tab lists everything other accounts have published. Search matches the name, the summary and the tags, so a dataset tagged chinese or support is findable without knowing who made it.

Each row shows the example count, file size, average tokens per row, and the licence. Open one to get the full card, the row viewer, and the measured structure — see The dataset page.

Your own published datasets are not repeated in Community; they have their own tab.

Using someone else's dataset

Open it and use Usage → launch a fine-tuning job, or pick it from the dataset dropdown in the fine-tuning wizard.

Two things to check before you spend GPU time on data you did not create:

  1. The licence. It is the one thing you cannot work out by looking at the data, and the one you may have to answer for later.
  2. The "Limitations" section of the card. A dataset that is entirely English, or entirely about refunds, will train a model that is entirely English or entirely about refunds.

Publishing your own

Use Publish a dataset on the registry page. It is the only place publishing happens, so there is one place to look when something is not listed.

The wizard has three steps.

1. Dataset

Pick from your datasets that have passed validation. Anything still uploading, validating or rejected is not offered — a dataset nobody can evaluate is not one worth listing.

Selecting one shows its example count, tokens per row and longest conversation, so a wrong pick is obvious before you write anything.

2. About

FieldWhy it matters
LicenceRequired. Pick from the common data licences, or enter your own.
SummaryA few lines, shown wherever the dataset is listed.
LanguageStored as a tag, so a Chinese dataset is filterable rather than buried.
TagsHow anyone finds this. Lower-cased and de-duplicated for you.

3. Card

The dataset card — the part that decides whether anyone actually uses what you published. It is written section by section, each with a prompt about what belongs there:

SectionWhat to write
What this isWhat the data teaches and who it is for
Where it came fromOrigin, date range, how it was gathered
How it was preparedFiltering, de-duplication, redaction — and what you dropped
Notes on structureOnly conventions a reader cannot see for themselves
Limitations and what it is not forKnown gaps, skews, and uses this data will not survive
Recommended useSettings that worked for you
Licensing and attributionWho owns the underlying data, and what a consumer must do

You do not describe the fields or the statistics — those are measured from your data and shown automatically.

Every section is optional, and the card can be finished later from the dataset's own page. A Markdown tab lets you write the whole thing directly if you prefer; switching back keeps anything written under headings of your own.

The section people skip is the one that saves the most time

"Limitations and what it is not for" is the section that stops someone spending an afternoon and a GPU discovering your data does not cover their case.

What publishing does and does not do

Does: make the dataset readable by every account, and list it in the registry.

Does not: copy your file anywhere, or change anything about jobs already running.

Withdrawing

Withdraw on the Published by you tab makes a dataset private again. It stops appearing in the registry immediately.

Withdrawing does not delete the dataset, and it does not reach into accounts that already copied it — a copy taken while it was published is theirs. Treat publishing as a decision you can stop, not one you can undo.

Storage

Published datasets count against your account's storage allowance like any other. Datasets are not billed separately today; the allowance exists so an accident cannot run away. See Datasets → Limits.

Next steps

  • Datasets — creating and validating one in the first place
  • Launching a job — training against a dataset, yours or someone else's