Retention
How long we keep your data and your models, and what stays after that.
The record of each dataset and model stays: its name, its numbers, the verdict against Jev and the audit of that verdict. What can be downloaded or trained on is kept for a limited time.
| What | How long | Then |
|---|---|---|
| Original upload | Until it is masked | Deleted right after masking. Only the masked copy is ever used. |
| A model's files (ONNX, weights, package, test examples) | 30 days after the training ends | Deleted. The record and the audit stay. |
| A model that did not beat Jev | At once | Its files are deleted when the verdict comes out: the credit came back, there is nothing to deliver. |
| A published model | Until it is on your Hugging Face | Its files leave our storage; your repository is where it lives and is downloaded from. |
| A dataset's data | 90 days after its last use | Deleted, including the copies we keep of previews and samples. Starting a training on it, or copying it to adjust, counts as a use and adds 90 days. |
| An upload that never finished, or failed | 30 days | Deleted with its record. |
| A closed account | 30 days after you ask | Everything is deleted. Until then you can cancel, and your API keys stop at once. |
The dashboard shows the dates: Available until on Models and Data. The API returns them on every model
(files_expire_at, and files_deleted_at once they went) and dataset (data_expire_at, data_deleted_at).
After expiry
- Downloading an expired model answers
410with"error": "expired". Its page, metrics and audit stay. - Expired data cannot start a training, be structured, labeled or copied (
410). To train again, upload it again. - Download your model within 30 days of its training, or publish it to your Hugging Face from its page.
Payment records are kept for 5 years and access logs for 6 months, as Brazilian law requires.