Standard / soft law
General-purpose AI models should maintain downstream documentation
Providers (signatories) of general-purpose AI models commit to drawing up and maintaining model documentation — via the AI Office's Model Documentation Form — covering technical specifications, training-data information, and energy consumption per AI Act Art. 53(1)(a)-(b) and Annexes XI/XII, and making it available to the AI Office and to downstream providers integrating the model. Detect a GPAI model release with no downstream/AI-Office documentation artifact.
Who it applies to
- Duty falls on: provider
- Systems covered: gpai
- Providers of general-purpose AI models placed on the EU market (Transparency + Copyright chapters apply to all GPAI providers under Art. 53). GPAI obligations apply from 2025-08-02; models released before then have until 2027-08-02. Signing the Code is voluntary but demonstrates AI Act compliance.
The guard to add
Organizational artifact to keep (not verifiable from code); the guard is the record, its owner and its upkeep.
Maintain versioned model documentation for each general-purpose model covering intended uses, training, and limitations, and publish it where downstream providers get the model.
Model documentation for each general-purpose model version (a model card or documentation page, such as model_card.md or the Hugging Face model card README) written for downstream integrators: intended and out-of-scope uses, architecture and technical specifications, how and on what data the model was trained, evaluation results, and known limitations. The provider's documentation owner revises it with each released version and keeps earlier versions retrievable. The repository holds its source, and the release pipeline publishes it alongside the weights or API so integrators receive it at the point of integration.
Where it goes: 12 repository artifacts, 14 user-facing text, 11 CI/CD pipeline.
What this provision adds:
- Draw up the documentation with the AI Office's Model Documentation Form, covering technical specifications, training-data information, and energy consumption.
- Make the documentation available to the AI Office and to downstream providers integrating the model.
Example (Hugging Face Hub release script), before:
from huggingface_hub import HfApi
HfApi().upload_folder(folder_path='out/', repo_id='org/base-7b')After:
from huggingface_hub import HfApi, ModelCard
card = ModelCard.load('out/README.md')
required = ['## Uses', '## Training Details', '## Bias, Risks, and Limitations', '## Technical Specifications']
missing = [s for s in required if s not in card.text]
if missing:
raise SystemExit(f'model card incomplete: {missing}')
HfApi().upload_folder(folder_path='out/', repo_id='org/base-7b')Control: General-purpose model without downstream documentation. The same guard addresses 1 item. Engineering guidance, not legal advice.
Rule id eu-gpai-cop.model-documentation · review status: primary source derived