Dataset policy

Rights you must have

Before you upload a dataset, you declare for each file:

• where the data came from;
• the license or agreement it is held under;
• that you have the right to use it to train and evaluate an adapter, and, if you publish, to distribute that adapter;
• whether it contains personal data, and if so, your lawful basis for processing it.

Data scraped against a site's terms, customer data you aren't permitted to use for training, and material under terms that forbid model training are not acceptable.

What we check

Uploads are validated before any training runs. We reject:

• malformed conversations or rows that don't match the expected schema;
• instructions aimed at the training or evaluation infrastructure;
• credentials, API keys and other secrets;
• illegal material.

Automated checks miss things. Passing them does not move responsibility for the data from you to us.

Private by default

• Datasets are private to your organization. They are never published and never shown to people who license your patch.
• Datasets are never placed on a blockchain, in whole or in part. Only digests of packages, manifests and terms are registered on chain.
• Training and evaluation run in isolated workers. Uploaded data is treated as data and is never executed.
• Sealed held-out cases are visible only to authorized reviewers, so a benchmark can't be tuned against.

A published patch's evidence includes the digests of the datasets and suites it used, plus aggregate scores. Digests identify data without revealing it. Example cases are shown publicly only from splits you mark as public.

What an adapter can reveal

Model weights trained on data can sometimes reproduce parts of it. If you publish a patch, assume that a determined person with the adapter may be able to extract fragments of the training data. Don't train a patch you intend to publish on data you couldn't accept leaking.

Retention and deletion

• You can delete a dataset at any time. Deletion removes the stored file and derived training splits, subject to backup expiry.
• Records of the dataset's digest, row count and rights declaration are kept with any evaluation or version that used it, so the evidence stays auditable.
• Deleting a dataset does not delete adapters already trained on it, and it does not recall published versions that others have downloaded.

First-party data

The first Retoque patch is trained and evaluated on synthetic data generated from templates, payee names and amounts written by Retoque. It contains no personal or customer data.

Complaints

If you believe a dataset or patch uses your data without permission, report it with as much detail as you can. We may suspend the related listing while we review. A published contact address will be added before paid licenses go live.