Audit modelModel Documentation Form (GPAI Code of Practice - Transparency Chapter)

The unique identifier for the model (e.g. Llama 3.1-405B), including the identifier for the collection of models where applicable, and a list of the names of the publicly available versions of the concerned model covered by the Model Documentation.

Evidence that establishes the provenance and authenticity of the model (e.g. a secure hash if binaries are distributed, or the URL endpoint in the case of a service), where available.

Date when the model was first released through any distribution channel.

Date when the model was placed on the Union market.

If the model is the result of a modification or fine-tuning of one or more general-purpose AI models previously placed on the market, list the model name(s) (and relevant version(s) if more than one version has been placed on the market) of those model(s). Otherwise write ‘N/A’.

A general description of the model architecture, e.g. a transformer architecture [Recommended 20 words]

A general description of the key design specifications of the model, including rationale and assumptions made, to provide basic insight into how the model was designed. If any other please specify. [Recommended 100 words]

The total number of parameters of the model, recorded with at least two significant figures, e.g. 7.3*10^10 parameters.

Provide a link to the acceptable use policy applicable or indicate that none exists

A description of either :

(i) the uses that are intended by the provider (e.g. productivity enhancement, translation, creative content generation, data analysis, data visualisation, programming assistance, scheduling, customer support, variety of natural language tasks, etc..) or

(ii) the uses that are restricted and/or prohibited by the provider (beyond those prohibited by EU or international law, including Article 5 AI Act), in both cases as specified in the information supplied by the provider in the instructions for use, terms and conditions, promotional or sales materials and statements, as well as in the technical documentation.

If specifying (i) or (ii) is incompatible with the nature of the license under which the model is provided, then 'N/A' can be entered. [Recommended 200 words]

A list or description of either

(i) the type and nature of AI systems into which the general-purpose AI model can be integrated or

(ii) the type and nature of AI systems into which the general-purpose AI model should not be integrated.

Examples may include autonomous systems, conversational assistants, decision support systems, creative AI systems, predictive systems, cybersecurity, surveillance, or human-AI collaboration. [Recommended up to 300 words]

A general description of the technical means (e.g. instructions for use, infrastructure, tools) required for the general-purpose AI model to be integrated into AI systems. [Recommended 100 words]

A description of any hardware, including the version, required to use the model, where applicable. If not applicable (e.g. model offered via an API), 'N/A' should be entered. [Recommended 100 words]

A description of any software, including the version, required to use the model where applicable. If not applicable, 'N/A' should be entered. [Recommended 100 words]

A general description of the main steps or stages involved in the training process, including training methodologies and techniques, the key design choices, assumptions made and what the model is designed to optimise for, and the relevance of different parameters, as applicable.

For example, “the model is initialized with randomly selected weights and optimised using gradient-based optimization via the Adam optimizer in two steps. First, the model is trained to predict the next word on a large pretraining corpus using the cross-entropy loss, passing over the data for a single epoch. Second, the model is post-trained on a dataset of human preferences for 10 epochs to align the model with human values and make it more useful in responding to user prompts“.

[Recommended 400 words]

A description of how and why key design choices were made in model training. [Recommended 200 words]

A description of the methods used to obtain and select training, testing, and validation data, including methods and resources used to annotate data, and models and methods used to generate synthetic data where applicable.

For data previously obtained from third parties, a description of how the provider obtained the rights to the data if not already disclosed in the public summary of training data published in accordance with Article 53(1), point (d).

[Recommended 300 words]

The size (in number of data points) of the training, testing, and validation data respectively, together with the definition of the unit of data points (e.g. tokens or documents, images, hours of video or frames), recorded with at least one significant figure (e.g. 3x10^13 tokens).

The size (in number of data points) of the training, testing, and validation data respectively, together with the definition of the unit of data points (e.g. tokens or documents, images, hours of video or frames), recorded with at least two significant figures (e.g. 1.5x10^13 tokens).

A general description of the scope and main characteristics of the training, testing and validation data, such as domain (e.g. healthcare, science, law,...), geography (e.g. global, restricted to a certain region,...), language, modality coverage, where applicable. [Recommended 200 words]

General description of the data processing involved in transforming the acquired data into training, testing, and validation data for the model, such as cleaning (e.g. filtering out irrelevant content such as advertisements), normalisation (e.g. tokenizing), augmentation (e.g. back-translation). [Recommended 300 words]

A description of any methods implemented in data acquisition or processing, if any, to detect the presence of unsuitable data sources considering the model’s intended uses, including but not limited to illegal content, child sexual abuse material (CSAM), non-consensual intimate imagery (NCII), and personal data leading to its unlawful processing. [Recommended 400 words]

A description of any methods implemented in data acquisition or processing, if any, to address the prevalence of identifiable biases in the training data. [Recommended 200 words]

A description of what period is being measured along with the range that its duration falls under, within the following ranges: less than 1 month, 1—3 months, 3—6 months, more than 6 months.

A description of what period is being measured along with the duration in wall clock days (e.g. 9x10^1 days) and in hardware days (e.g. 4x10^5 Nvidia A100 days and 2x10^5 Nvidia H100 days), both recorded with at least one significant figure.

Reported in floating point operations and recorded up to its order of magnitude (e.g. 10^24 floating point operations).

Reported in computational operations and recorded with at least two significant figures (e.g. 2.4x10^25 floating point operations).

In the absence of a delegated act adopted to detail measurement and calculation methodologies, describe the methodology used to measure or estimate the amount of computation used for training.

Measured or estimated amount of energy used for training, reported in Megawatt-hours and recorded with at least two significant figures (e.g. 1.0x102 MWh). If the amount of energy used for training cannot be estimated due to the lack of critical information from a compute or hardware provider, enter ‘N/A’.

In the absence of a delegated act adopted to detail measurement and calculation methodologies, describe the methodology used to measure or estimate the amount of energy used for training.

Where the energy consumption of the model is unknown, the energy consumption may be estimated based on information about computational resources used.

If the amount of energy used for training cannot be estimated due to a lack of critical information from a compute or hardware provider, the provider should disclose the type of information they lack.

[Recommended 100 words]

Reported in floating point operations, recorded with at least two significant figures (e.g. 5.1x10^17 floating point operations).

This item relates to energy consumption during inference, which makes up the “energy consumption of the model” (Annex XI, 2(e), AI Act) together with energy consumption during training. Since energy consumption during inference depends on more than just the model itself, the information required for this item is limited to relevant information depending only on the model, namely computational resources used for inference.

In the absence of a delegated act adopted to detail measurement and calculation methodologies, provide a description of a computational task (e.g. generating 100000 tokens) and the hardware (e.g. 64 Nvidia A100s) used to measure or estimate the amount of computation used for inference.