compar:IA: The French Government's LLM arena to collect French-language human prompts and preference data

The French government, through the DINUM and Etalab departments, officially unveiled the 'compar:IA' platform in Paris this February to address the critical shortage of high-quality French-language training data for Large Language Models (LLMs). This initiative functions as a public leaderboard and evaluation arena where human users interact with various AI models to provide preference data, specifically designed to counter the English-centric bias currently dominating the global artificial intelligence landscape. By launching this open-access tool, France aims to enhance the cultural alignment, linguistic nuance, and safety protocols of AI systems operating within the Francophone world. Technological development in the AI sector has historically been hampered by a lack of diverse datasets, with methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) relying heavily on English-language inputs. The 'compar:IA' project seeks to mitigate these issues by collecting human prompts and preferences in an open-source framework, allowing researchers to see how models handle specific French cultural contexts and idiomatic expressions. This move is seen as a strategic effort to ensure that sovereign AI development remains competitive and representative of the French national identity and values. Beyond simple translation, the platform focuses on solving the 'reduced performance' often seen when global models are applied to non-English tasks. By gathering authentic human interaction data, the French government intends to bridge the gap in safety robustness and cultural accuracy that frequently plagues systems pre-trained primarily on American or British web data. The resulting datasets are expected to be made public, providing a rare resource for developers who previously lacked access to large-scale, non-proprietary preference data for the French language.