Dataset Datasheet

How to use this document

• What it is for. This is a base datasheet to document each dataset used to train, fine-tune, validate, test or operate an AI system, including knowledge bases for retrieval-augmented generation (RAG) systems and evaluation sets. It records origin, usage rights, contents, personal data, quality, known biases, security, retention and permitted uses.
• Who should own it. The dataset owner, usually in the data team or in the business function that generates the data, with review by privacy and, when third-party licenses are involved, by legal. Enter the function in [RESPONSIBLE FUNCTION] and the person in [OWNER].
• Decisions to take first. Decide from which use a datasheet is required (Section 1.2), how much detail to ask for low-risk datasets (Section 2.4) and where datasheets are stored so they are linked to the inventory under the "AI Model and Dataset Inventory Policy".
• Fields to complete. Complete the "Content" column of each block in Section 3. The "Example" column shows a complete fictitious case to guide the level of detail; delete it in your working version if you prefer. Use [COMPLETE] only as a temporary marker: a datasheet with pending fields is not approved.
• Country variants. This document has no country variants. If the dataset contains personal data from several countries, check in "Related regulatory references" the legal basis, purpose, database registration and retention requirements of each and reflect them in block D.
• Review. Privacy reviews block D; legal reviews block B when there are licenses, contracts or data obtained from the internet; the model owner reviews blocks E and H.
• Approval. The dataset owner approves the datasheet, with the reviews listed, before the dataset is used. Datasheets are internal documents; they can be shared with auditors, customers or authorities when requested.

Dataset Datasheet

1. Purpose and when it is completed

1.1 This datasheet documents a dataset used by [ORGANIZATION NAME] in an AI system, so that any authorized person can know what it contains, where it comes from, under what rights it is used, what limitations it has and what it may and may not be used for.

1.2 One datasheet is completed for each relevant version of a dataset used to:

a) train or fine-tune a model;
b) validate or test a model before deployment or after a change;
c) feed an AI system in operation as a knowledge base or reference source;
d) periodically evaluate the performance or bias of an AI system.

1.3 The datasheet is completed before the dataset is used and updated when its contents, origin, terms of use or the systems using it change. Each new version of the dataset is recorded in Section 4.

2. General instructions

2.1 Write in plain language, understandable to someone who did not take part in creating the dataset.

2.2 When a field does not apply, write "Not applicable" and the reason. A blank field is treated as pending.

2.3 Statements on quality, representativeness or bias are backed by documented analyses, linked in the relevant field.

2.4 [OPTIONAL] Low-risk datasets (no personal data, no restrictive licenses and used only in Low-risk systems) may be documented with blocks A, B, C, G and H only.

2.5 If the dataset is supplied by a third party, the information in blocks B, C, D and E is requested from it and its answer is filed as evidence.