Connect
Sign up to receive our monthly newsletter.
Inquiries
Have a question? Contact us here.
Job Opportunities
Learn about job openings.
Digital Public Goods Alliance - 2026
Sign up to receive our monthly newsletter.
Have a question? Contact us here.
Learn about job openings.
DPGs for AI Collection
All DPGs included in the DPGs for AI Collection must meet the requirements outlined in the DPG4AI collection criteria.
The information below reflects the self-reported responses provided as part of the assessment process.
Solution Type: Open Source Software
DPG Compliance & Profile Page: https://www.digitalpublicgoods.net/r/comparia
Description: Compar:IA is a conversational AI arena where users can explore and compare over 30 different models. Its mission is to promote awareness of model pluralism, bias, and environmental impact, while simultaneously building one of the world’s largest non-english alignment datasets.
Assessment Status: Under Review
Review Date: 2026-06-14
Category Fit:
Open Source Software (Apache-2.0; Python/Gradio/Svelte). Compar:IA is a public 'LLM arena' web app from the French Ministry of Culture. Lifecycle roles: Verification & Validation (blind side-by-side model comparison, preference voting, leaderboard, bias/energy surfacing) and Data Acquisition & Preparation (it builds and publishes large non-English human-preference / alignment datasets). Repo: https://github.com/betagouv/ComparIA
AI Lifecycle Utility:
Documented Relevance/ Impact:
Live government service opened Oct 2024; >600k free-form prompts and >250k preference votes (snapshot Feb 2026). Publishes three open datasets — conversations, votes, reactions — on Hugging Face to improve non-English (esp. French) model alignment. Paper: arXiv:2602.06669. Live in French & Danish; Swedish/Estonian/Lithuanian planned. https://huggingface.co/datasets/ministere-culture/comparia-conversations
Adoption Readiness Level:
L4 Orchestrated / Optimized
Adoption Readiness Evidence:
Production government service (comparia.beta.gouv.fr) under beta.gouv.fr; open repo with README and active development; datasets versioned on Hugging Face (ministere-culture/comparia-conversations, -reactions). Research paper documents methodology and dataset releases.
Interoperability Level:
L4 Integrated — versioned API/SDK, plug-in capable, automated synchronization
Interoperability Evidence:
Provider-agnostic via litellm (OpenAI/Anthropic/Vertex/Mistral/xAI/OpenRouter plus open models Qwen/DeepSeek/Aya/Phi/OLMo). Programmatic dataset export (API/JSON/CSV/Markdown). Standards: WCAG/RGAA, DSFR, HTTPS/TLS, JSON, YAML, UTF-8. Export utility: https://github.com/betagouv/ComparIA/blob/develop/utils/export\_dataset.py
Responsible Practices Level:
L3 Fully Documented — comprehensive limitations, biases, and failure modes statement
Equity & Inclusion:
Responsible AI Tooling:
Responsible Practices:
Inclusion & Autonomy:
Responsible Practices Evidence:
Core mission is surfacing model limitations — cultural/linguistic bias, model pluralism, and per-response environmental (energy) impact — documented in the arXiv paper (2602.06669) and on-site (donnees-personnelles, modalites). PII handling: explicit consent + a-posteriori PII scanning/exclusion before any dataset is published.
