Data for AI: datasets, corpora, evaluations and red teaming
Datasets to use as they are: train a model, feed a RAG, test it or attack it.
AI dataset on the population of Dutch municipalities
41 dialogues with the CBS rows in context (Dutch municipalities, 2025): the model reads the figure, cites it, or declines.
AI dataset on the population of Irish counties
39 dialogues with the CSO rows in context (26 counties, census 2022): the model reads the figure, cites it, or declines.
UK dataset for anonymising personal data (UK GDPR)
20 examples to anonymise UK texts (UK GDPR): NHS numbers, +44 phones, GB IBANs, number plates, postcodes. Fictional people and numbers. Includes cases with no…
AI dataset on the population of U.S. states
43 dialogues for an AI assistant on U.S. state populations: 2025 estimates, change since 2020, regions (Census Bureau). Cites its source.
AI dataset: finding the HS customs code of a product
43 dialogues with an extract of the Harmonized System in context: the model gives the HS code, asks for a detail or declines.
Corpus: AI resource definitions (English)
Definitions of every AI resource type on Contexte Tech, one chunk each with the guide URL: a small English corpus ready for RAG.
AI dataset: finding the NACE code of an activity
44 dialogues with an extract of the NACE Rev. 2.1 classification in context: the model gives the code, asks for a detail or declines.
AI dataset: agent finding the HS customs code with a tool
43 agent traces: call to the nomenclature search tool, then the HS code, a follow-up question or a refusal.
AI dataset (DPO): HS customs code without guessing
40 preference pairs: HS code read in the nomenclature and cited, versus a neighbouring or remembered code.
AI evaluation of English text anonymization
Checks that an anonymization model or context lets no U.S. personal data through (email, phone, SSN, date of birth, medical record number).
Red teaming: 12 attacks on a customer service chatbot
12 attacks to replay on a U.S. e-commerce support chatbot: impersonation, data leaks, prompt injection, unauthorized refunds. Expected reply for each.
Red teaming: 10 tests on a U.S. HR assistant
10 tests for an internal HR assistant in the U.S.: employee data, hiring discrimination (EEOC), impersonation, injection, and 2 legitimate questions.
AI dataset: agent finding the ESCO occupation with a tool
60 agent traces: call to the ESCO search tool, occupations returned, then occupation and code or a refusal.
AI dataset: finding the ESCO occupation of a job title
60 dialogues with ESCO occupations in context: the model gives the occupation and its code, asks for a detail or declines.
AI dataset (DPO): ESCO occupation without guessing
57 preference pairs: occupation read in the ESCO classification and cited, versus a neighbouring or remembered one.
AI dataset: normalizing US addresses
36 dialogues with U.S. Census geocoder lines in context: normalized address, a follow-up question or a refusal.
What is data for AI?
Data, here, is a set of examples or documents used as it is: to train a model (dataset), give it knowledge in a RAG (corpus), measure its answers (evaluation) or attack it to find its weaknesses (red teaming). It does not tell the model what to do: it shows it cases, real or fictional, with their source.