# Start a dataset registry

What each dataset is, where it came from, its license, its versions and what changed between them. Legal owns the License field, and pipelines register new versions themselves.

## An example

**You:** Register dataset v8: v7 plus 40B tokens of filtered web data from crawl 2026-09, dedup by MinHash. License mix: CC-BY and permissive code. Owner Mira.

**ADoor:** Added 1 entry to Datasets

**Datasets**

| Dataset | Version | Change | License |
| --- | --- | --- | --- |
| pretrain-mix | v8 | +40B web, MinHash dedup | CC-BY, permissive |
| pretrain-mix | v7 | +code subset | CC-BY, permissive |

## What you can ask later

- “Which datasets can we use commercially?”
- “What changed between v6 and v8?”
- “Which models were trained on data that includes the 2026-09 crawl?”
- “Who owns the instruction-tuning data?”

## Who can see what

- **Data owners** register and edit their own datasets (one grant per person).
- **Everyone on the team** reads the registry.
- **Legal** edits the License field, which nobody else can change.
- **A pipeline job** registers new versions with short-lived CI access.

## How it compares with Hugging Face Hub or a data catalog

The Hugging Face Hub is better for hosting and sharing the data itself, with dataset cards and versioned files. Data catalogs are better at automatic lineage from your warehouse.

What ADoor adds is real access control on a lightweight registry: legal owns the license field, pipelines register versions, and SQL joins datasets to the runs that used them.

## How to start

1. Sign in to ADoor and connect an assistant like ChatGPT. Claude, Grok and Cursor connect the same way, through MCP.
2. Tell it: “Set up a dataset registry: Datasets (name, version, description, source, license, owner, size, created), linked to Runs.”
3. Give legal edit access to License only, and your pipeline short-lived CI access to add versions.

[Get your account](https://app.adoor.halmoslabs.com/signup)

## Questions people ask

### Does ADoor store the data files?

No. Store the files where you train, and keep their location here.

### Can I link datasets to runs?

Yes. Link each run to the dataset version it used.

### How much does ADoor cost?

The free plan holds 3 tables and 100 rows, enough to start a list like this. Pro is $10 a month per workspace, with 1 GB of storage and no per-GB charges.

### Which assistants work with ADoor?

ChatGPT, Claude, Grok and Cursor connect through MCP. You sign in once per assistant, approve it, and choose what it can reach.

## Related

- [Keep a model training log](/use/model-training-log/)
- [Start a labeling and annotation queue](/use/annotation-queue/)
- [Track eval results over time](/use/eval-results/)
- [ADoor vs Supabase](/compare/adoor-vs-supabase/)
- [ADoor vs Airtable](/compare/adoor-vs-airtable/)
- [All use cases](/use/)

Updated October 20, 2026.