Glossary

Terms as used on this site.

Terms

Term	Definition
Pretraining data	Large, broad corpora used to train a base model.
Post-training	Later stages (SFT, preference tuning, RL) that shape behaviour with smaller, targeted data.
SFT	Supervised fine-tuning on example inputs and desired outputs.
RLHF	Reinforcement learning from human feedback, typically using preference judgements.
Evaluation set	Held-out data used to measure a model; training leakage can compromise the validity of the evaluation.
Chain of title	Documented sequence showing how rights passed to the current owner.
Field of use	The purposes a licence permits.
Datasheet	Structured documentation of a dataset's origin, composition and limits.
De-identification	Removing or transforming identifiers; not always equivalent to anonymisation under law.

Pretraining data: Large, broad corpora used to train a base model.
Post-training: Later stages (SFT, preference tuning, RL) that shape behaviour with smaller, targeted data.
SFT: Supervised fine-tuning on example inputs and desired outputs.
RLHF: Reinforcement learning from human feedback, typically using preference judgements.
Evaluation set: Held-out data used to measure a model; training leakage can compromise the validity of the evaluation.
Chain of title: Documented sequence showing how rights passed to the current owner.
Field of use: The purposes a licence permits.
Datasheet: Structured documentation of a dataset's origin, composition and limits.
De-identification: Removing or transforming identifiers; not always equivalent to anonymisation under law.

Tell us what you own, which rights you can license and the commercial scope you have in mind.