LLaMandement — France's AI Tool for Processing Parliamentary Budget Amendments

Each autumn, France's Parliament tables thousands of amendments to the annual Finance Bill, which Direction Générale des Finances Publiques (DGFiP) officials must read, route to the right ministerial expert team, and summarise for negotiators — an exercise nicknamed 'la nuit des amendements' (the night of the amendments).

To automate the classification/routing and drafting of neutral summaries of parliamentary budget amendments.

In 2023-24, DGFiP's data science team, working with DINUM and DILA, built LLaMandement by adapting Meta's open-source Llama 2 model through low-rank adaptation (LoRA) fine-tuning on more than 15,000 real amendment-and-summary pairs drawn from SIGNALE, the interministerial platform for managing legislative amendments. The method was published in a peer-reviewable arXiv paper (January 2024) and a version of the model and part of the training data were released openly on Hugging Face ('AgentPublic/LlaMAndement-7b').

DGFiP and Senate officials have cited a fall in per-amendment processing time from 6-10 hours of manual work to around 15 minutes. The authors reported summarisation quality 'close to that of human writers' on the specialised task, though without publishing a precise quantitative accuracy figure against a human baseline.

When the Senate considered adopting a version of the tool for its own 2024 budget session, budget rapporteur Jean-François Husson cautioned it 'will not significantly lighten' senators' real workload since freed-up time would likely go toward deeper amendment scrutiny; other senators raised digital-sovereignty concerns about building sensitive legislative infrastructure on an American foundation model. As of the cited reporting, a dedicated Senate deployment had not been finally decided, and use of the tool remains centred on DGFiP's own internal process.