GlossaryFloor 1 · The Modela solid block on its own: the prediction machineFloor 1 · The Model
training data
No. 050 · v2026-08FR: données d’entraînementTraining data is the set of examples from which a model was formed: what it can say and what it is ignorant of both depend on their content. Like the library where someone learned everything, and whose gaps become their own.
What it is not
Training data is not what you give the model when you talk to it: that is called the context, it holds for one conversation, and it trains nothing. Nor is it stored inside the model, which keeps only parameters. And it is not a set of interchangeable raw materials: its composition determines which languages are well served, which domains are mastered, the date beyond which the model knows nothing, and the biases it will reproduce.
In depth
Composition
Composition explains behaviour better than any other variable. An under-represented language gives poorer answers and a more expensive breakdown into tokens. An absent domain produces answers that are plausible but hollow, which is more dangerous than a refusal. An over-representation of certain points of view is found again in the phrasing. And since the corpus stops at a date, the model knows nothing of what came after, without knowing it: it will answer on those subjects with the same assurance as on the rest.
Sorting the corpus
Quality now takes precedence over quantity. Recent progress comes less from adding text than from choosing it better: filtering duplicates, removing poor quality content, balancing domains, and above all adding data carefully written to teach a behaviour rather than a body of knowledge. It is this last point that separates a raw model, which merely continues a text, from a usable model, which follows an instruction and refuses certain requests.
Legal status
The legal status of this material is the open subject of the period. Part of the corpora comes from the web, and therefore from protected works, and the disputes bear on the legitimacy of that use, on the remuneration of rights holders and on the presence of personal data. This has direct consequences for an organisation: the provenance of a model’s data is part of what gets assessed before adopting it, on the same footing as its performance, because it carries a risk that no technical measure reduces after the fact.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
You paste an internal report into a conversation. Does that train the model?
Level 2 · Distinguish
A model answers fluently but wrongly on a highly specialised subject. The most likely explanation?
Level 2 · Distinguish
Why does the provenance of training data concern an organisation that merely uses a model?
Who works with this 3 roles
The roles for which this term is part of the ordinary work.
Lexigraph, "Training data", v2026-08, https://www.lexigraph.org/en/training-data/, CC BY 4.0.