ANNEX XI
Technical documentation referred to in Article 53(1)(a) - technical documentation for general purpose AI model providers
Section 1
Information to be provided by all general purpose AI model providers
The technical documentation referred to in Article 53(1)(a) shall contain at least the following information, according to the size and risk profile of the model:
|
1.
|
A general description of the AI model for general purposes, including:
|
a)
|
the tasks to be performed by the model and the type and nature of AI systems into which it can be integrated;
|
|
b)
|
The "acceptable use policies" that apply;
|
|
c)
|
the release date and method of distribution;
|
|
d)
|
architecture and number of parameters;
|
|
e)
|
the modality (e.g., text, image) and format of the input and output;
|
|
|
2.
|
A detailed description of the elements of the model referred to in point 1 and relevant information on the development process, including the following elements:
|
a)
|
The technical resources (e.g., user instructions, infrastructure, tools) needed to integrate the general purpose AI model into AI systems;
|
|
b)
|
the specifications of the design of the model and training process, including training methods and techniques, the main design choices, including the rationale and assumptions made; for which optimization the model is designed and the relevance of the different parameters, as appropriate;
|
|
c)
|
Information on the data used for training, testing and validation, if applicable, including the type and origin of the data and curation methods (e.g., cleaning, filtering, etc.), the number of data points, their range and key characteristics; how the data were obtained and selected, as well as any other measures to detect the inappropriateness of data sources, and methods for detecting identifiable biases, if applicable;
|
|
d)
|
The computational tools used to train the model (e.g., number of floating point operations), duration, and other relevant details about the training;
|
|
e)
|
known or estimated energy consumption of the model.
|
Regarding point (e), if the energy consumption of the model is unknown, the energy consumption may be based on information about the calculation tools used.
|
Section 2
Additional information to be provided by systemic risk AI model providers
|
|
1.
|
A detailed description of evaluation strategies, including evaluation results, based on available public evaluation protocols and tools or other evaluation methods. Evaluation strategies include evaluation criteria, benchmarks and the methods for identifying limitations.
|
|
|
2.
|
If applicable, a detailed description of the measures taken to conduct internal and/or external testing aimed at discovering vulnerabilities (e.g., red teaming), model modifications, including tuning and refinement.
|
|
|
3.
|
If applicable, a detailed description of the system architecture with an explanation of how software components build on each other or provide information to each other and how integration into overall processing occurs.
|
