Model Card

How to use this document

• What it is for. This is a base datasheet that documents an AI model in a standard way: what it is for and what it is not for, which data it was built with, how it was evaluated, which biases and limitations it has, how it should be overseen and who is accountable for it. It is a shared reference for operations, risk, audit and the business and, once adapted, for customers who receive the model.
• Who should own it. The data science team that builds or fine-tunes the model drafts the card; the system owner approves it and keeps it current. Enter the function in [RESPONSIBLE FUNCTION] and the accountable person in [OWNER].
• Decisions to take first. Decide the level of detail each risk level calls for (aid B), which metrics and thresholds are set in Section 5, and whether there will be a summary version for customers or users (aid F).
• Fields to complete. One card per model and per relevant version. Replace the bracketed text and write "Not applicable" with a short reason instead of deleting fields. Keep [OPTIONAL] sections only if they apply, for example the generative AI section.
• Third-party models. If the model comes from a [VENDOR], complete the card with the documentation the vendor provides and note which information could not be obtained. That gap is a risk to record in Section 8. The provenance, license, restrictions by origin and hosting mode of a third-party model are documented in its "Third-Party and Open-Weight Model Provenance Sheet"; the card only cites its identifier.
• Country variants. This document has no country variants. If the model is a high-risk system, a general-purpose model or a regulated product (for example, medical software), use "Related regulatory references" to identify the additional technical documentation required and attach it to this card.
• Review and approval. The AI lead (Medium risk) or the [AI COMMITTEE] (High risk) reviews the card before go-live. Record the approval in Section 13 and in "Version history."

Model Card

Field	Value
Organization	[LEGAL ENTITY NAME]
Responsible function	[RESPONSIBLE FUNCTION]
Card owner	[OWNER]
AI system where it is used	[AI SYSTEM]
Date of this card version	[DATE]
Review frequency	[REVIEW FREQUENCY]
Information classification	[COMPLETE: internal, confidential or suitable for customers]

1. Model identification

Field	Content
Model name	[COMPLETE]
Version	[COMPLETE]
Model type	[COMPLETE: classification, regression, recommendation, vision, language processing, generative or other]
Origin	[COMPLETE: built in-house, fine-tuned on a third-party model or acquired from a vendor]
Base model and vendor, if applicable	[COMPLETE: name, version and [VENDOR]]
License and terms of use	[COMPLETE: summary; the detail for a third-party model is in its provenance sheet]
Provenance sheet identifier, if the model or its base model is third-party	[COMPLETE]
Weights hash of the deployed version, if hosted internally	[COMPLETE]
Status	[COMPLETE: Proposed, Under assessment, Pilot, In production, Under review or Retired]
Risk level (per the "AI System Inventory")	[COMPLETE: Low, Medium or High]
Identifier in the "AI System Inventory"	[COMPLETE]
System owner	[COMPLETE]
Development team	[COMPLETE]
Date of last validation	[DATE]
Contact for questions	[AI GOVERNANCE EMAIL]

1.1 If the model was fine-tuned or substantially modified from a third-party model, record here legal's assessment of the organization's role: that modification may change it under some laws (for example, art. 25 of Regulation (EU) 2024/1689). [COMPLETE]

2. Intended use

2.1 Purpose. [COMPLETE: problem the model solves and the output it produces].

2.2 Decisions it supports. [COMPLETE: which decision or process it supports and whether a person takes the final decision].

2.3 Intended users. [COMPLETE: who uses the outputs and what training they need].

2.4 Affected people. [COMPLETE: whom the outputs concern, including whether minors or vulnerable groups are involved].

2.5 Deployment context. [COMPLETE: channels, countries, languages, expected volume and systems it integrates with].

2.6 Out-of-scope uses. The model must not be used to:

• [COMPLETE: uses it was not designed or evaluated for];
• [COMPLETE: populations, languages or contexts not represented in the data];
• take decisions with legal or similarly significant effects on people without the human oversight described in Section 9;
• carry out any use prohibited by the responsible AI use policy of [ORGANIZATION NAME] or by law.