Enterprise dataset questions from AI labs

Complete records in a documented schema (whole conversations, tickets with their outcomes, code with history, documents with structure), typed pseudonyms in place of identifiers, a datasheet, a de-identification report and a rights record. The spec sheets list the fields per data type.

JSON Lines with one conversation per line, so a loader never has to stitch messages together. Parquet with the same field layout works for large deliveries. Documents are best as Markdown with a JSONL index.

Ask for leak-scan results, a human-reviewed sample with its error rate and recall per language, then run your own review on a random sample: search for contact details, read greetings and signatures, and check token consistency.

A datasheet covering composition, provenance, processing, de-identification, permitted use and AI Act inputs, plus file checksums, a code book for labels, and the rights record. The datasheet template on this site covers all of it.

The obligations sit with the model provider, but several answers in its documentation and public summary need supplier facts: token counts, source category, nature of the content, languages and screening measures. Collect them in the contract.

Only when exclusivity is worth its cost. Non-exclusive licences are the norm. An exclusive window makes sense for evaluation sets you need to keep uncontaminated, or for a capability you want to ship first.

Yes, when tasks have recorded outcomes that a grader can check. Support tickets with resolution codes and code changes with passing tests are the usual starting points.

Use the request form or email hello@aidataset.broker. Asking costs nothing and commits you to nothing. Datasets are assembled to spec, de-identified, documented and delivered under a licence with a datasheet.