Automating AI Model Documentation with the NVIDIA MCG Toolkit

As AI model complexity increases and regulations such as the EU AI Act become stricter, creating Model Cards to demonstrate model transparency and accountability has become essential. However, manual creation methods are not only time-consuming but also fail to keep pace with the speed of model updates, making it difficult to keep the data current.

The NVIDIA MCG (Model Card Generator) Toolkit is a containerized pipeline that reads data directly from source code to automatically generate Model Card++ format documentation in under 1 minute. This toolkit operates through modular stages of Ingestion $\rightarrow$ Extraction $\rightarrow$ Rendering.

The detailed process is as follows:

• Ingestion: Receives URLs or files from GitHub, GitLab, HuggingFace, etc., and classifies and processes them into documents, configuration files, and code.
• Extraction: Utilizes a RAG (Retrieval-Augmented Generation) pipeline based on NVIDIA NIM. NVIDIA Nemotron RAG performs high-precision embedding and reranking, while GPT-OSS-120B generates spec-compliant JSON data based on the extracted information.
• Rendering: Converts the structured JSON data into easy-to-read Markdown format through configurable templates.

The final output is completed with an overview along with 4 detailed cards covering Bias, Explainability, Privacy, and Safety & Security.