Glossary
AI training data, defined
The terms you’ll meet when licensing software engineering data for models.
- Pre-training data
- Large-scale text and code used to train a foundation model from scratch or to continue its pre-training. For code models this includes source repositories, documentation, and commit history.
- Fine-tuning data
- Smaller, curated datasets used to adapt a pre-trained model to a task or behaviour, such as following instructions or writing code in a house style.
- Evaluation data (benchmark)
- Held-out tasks used to measure model quality. Useful only if the model has not seen them during training — see benchmark contamination.
- Benchmark contamination
- When evaluation tasks or their solutions appear in training data, inflating reported scores. Contamination checks compare datasets against known public corpora.
- Provenance
- The documented origin of a dataset: who created it, who owns it, how it was collected, and what has happened to it since.
- Rights review
- An assessment of who owns the data, what licences and restrictions apply, and which uses (pre-training, fine-tuning, evaluation) are permitted. It records what was and was not reviewed.
- Permitted use
- A specific purpose a licence allows, such as pre-training, fine-tuning, or evaluation. Different uses can be licensed separately and with different exclusivity.
- Inbound licence
- The licence terms under which the supplier received the underlying material, for example an open-source licence on a dependency. Inbound terms can limit what the supplier may license onward.
- Dataset card
- A structured description of a dataset: contents, schema, collection method, checks performed, known limitations, and permitted uses.
- Secret scanning
- Automated detection of credentials such as API keys, tokens, and private keys in source code and history, so they can be removed before data is shared.
- PII scrubbing
- Removing or pseudonymising personal data — names, emails, and identifiers — from text such as commit messages, issues, and review comments.
- De-duplication
- Removing exact and near-duplicate records so a model does not over-weight repeated content and evaluation sets stay clean.
- Issue-to-fix history
- A linked record of a problem report, the code change that resolved it, the review discussion, and the test outcome. Valuable for training and evaluating coding agents.
- Coding agent
- An AI system that plans and carries out software engineering tasks — reading code, making edits, running tests — rather than only completing a single snippet.
- Data licensing
- A written agreement granting a buyer defined rights to use a dataset for stated purposes, for a term, with agreed restrictions and delivery terms.
- Exclusivity
- Whether a licence is the only one of its kind. A non-exclusive licence can be granted to several buyers; an exclusive one means no other buyer gets the same dataset version for the same uses while it runs. Which is possible for each dataset is set in its rights review.
- Sample
- A small, scrubbed, time-limited extract of a dataset, shared for evaluation once a request is approved and where the listing has one. It may not be used for training and is never published.
- Data partner
- An organisation that owns engineering data, or is authorised to license it, and lists it through ModelData Exchange. Also called the supplier.
Questions about your data?
Share what you need or what you own. We reply with next steps — or tell you honestly if we can’t help.