What are training, validation and test data?
They are three disjoint slices of your data, each with one job.
| Set | Used for | How often the team looks | What goes wrong if misused |
|---|---|---|---|
| Training | Fitting the model’s parameters | Constantly | Memorized noise, overfitting |
| Validation | Tuning settings, choosing a model, early stopping | Every experiment | Settings tuned to one lucky slice |
| Test | One final estimate of real performance | Once per decision | Scores that do not survive production |
A test set used to make decisions stops being a test set. The same rule holds for LLM eval datasets: when public test questions leak into training data, scores inflate, the problem called benchmark contamination.
What is an example of training data?
One row of input plus whatever the model should learn to produce.
| Task | One training example | Where the target comes from |
|---|---|---|
| Spam filter | An email’s text and sender | Users marking messages as spam |
| Invoice amount prediction | Line items, supplier, date | The amount actually paid |
| LLM pretraining | A passage of web text | The next token in the passage itself |
| Instruction fine-tuning | A prompt | A response written by a person |
The first two are supervised learning with labels people or systems recorded. Pretraining is self-supervised: the text is its own answer key, which is why its scale is enormous. Meta’s Llama 3 “is pretrained on over 15T tokens that were all collected from publicly available sources”, a dataset “seven times larger than that used for Llama 2” (Meta AI blog, 18 April 2024, read 16 November 2026).
What makes training data good?
It looks like the data the model will meet in production, and its labels are right. Coverage beats volume: a million examples from one region teach nothing about another. Labels need a written guideline and review, the work of data annotation, and the labels you trust become your ground truth. Gaps can be filled with synthetic data, validated against real cases.
For high-risk systems the EU AI Act makes this a legal duty. Article 10(3) requires training, validation and testing data sets to be “relevant, sufficiently representative, and to the best extent possible, free of errors and complete”, applying from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems (AI Act Explorer, Article 10, read 16 November 2026). Write down where every training set came from; you will be asked.
Do you need your own training data to use an LLM?
Rarely, because most LLM projects never train a model at all. Prompting and RAG use a pretrained model unchanged, and your documents reach it as context, not as training data. Your examples become training data only when you fine-tune. What every LLM project does need is evaluation data from its own domain. Model providers, meanwhile, owe transparency: Article 53(1)(d) requires general-purpose model providers to publish “a sufficiently detailed summary about the content used for training”, in force since 2 August 2025 (AI Act Explorer, Article 53, read 16 November 2026), using a Commission template published on 24 July 2025 (European Commission, read 16 November 2026).