# Evaluation dataset

URL: https://sellmydatatoai.net/glossary/evaluation-dataset

An evaluation dataset is a set of tasks with expected answers and scoring rules, used to measure how well a model performs rather than to train it.

An evaluation dataset is a collection of tasks, each with an expected answer or a scoring rule, used to measure a model's performance. It is kept separate from training data so that scores reflect real ability, not memory.

## Why are evaluation datasets valuable?

Evaluation datasets are valuable because public benchmarks leak into web-scale training data over time, so labs need fresh, private tests in specialist domains. A set built from real expert work, with answers checked by qualified reviewers, shows whether a model can do the job.

## What does a good evaluation dataset contain?

- Tasks drawn from real work, not invented examples
- Reference answers or rubrics written by experts
- Scoring rules that two reviewers would apply the same way
- Metadata on difficulty and domain
- Confirmation that the set has never been published

Sellers should license evaluation sets on terms that bar training on them, and buyers can [submit a data requirement](https://sellmydatatoai.net/request-data) for a specific domain. See [hold-out set](https://sellmydatatoai.net/glossary/hold-out-set) and [evaluation datasets](https://sellmydatatoai.net/buy-ai-training-data/evaluation-datasets).