I enjoyed contributing to the European AI Office on frontier LLMs, available here.

Nous co-organisons une journée “frugalité en TAL et ML” à Paris le 2 avril 2026. Cette journée est soutenue par les GDR TAL, IASIS et MADICS du CNRS. Nous aurons le plaisir d’accueillir Gaël Varoquaux, Julia Gusak, Nicolas Keriven et Raphaël Troncy pour les séminaires invités. Vous pouvez soumettre un résumé de vos travaux pour les présenter à la journée. Plus d’infos ici: https://ia.loria.fr/frugalday/

Mille-Pensées is one of the best 7b LLM that reasons in French that we have post-trained from a Qwen2.5. When you ask most top recent 7b LLMs to solve a maths question in French, most of them will correctly answer in French, but will reason in English. We have therefore built a post-training pipeline and data-mix that makes the LLM also reason in French, and applied this pipeline to produce the Mille-Pensées LLM.

The main event at the AI Action Summit organized by L’Elysée the 10th and 11th February 2025 at Grand Palais in Paris will host an NLP session co-organized by Nicholas Asher, Serena Villata and Christophe Cerisara. More details can be found on the dedicated web page.

ENACT is a Cluster-IA project that aims at developping several multimodal LLM for various application domains. For more info, see the project website.

OpenLLM-France is a BPI project lead by Linagora, which aim is to train really open LLMs that support well in particular the French language. For more info, see the project website.

LLM4All is an ANR project dedicated to finetuning LLMs to ensure that they will stay up-to-date. Two application domains are considered: meetings understanding, and emergency calls for hospitals. For more info, see the project website.

We are developping inference and training scripts, with the help of CNRS GENCI AI engineers, to deploy on the Jean Zay cluster to help the community starting on Jean Zay with giga-models. For more info, see our website, our PLM4All gitter and our PLM4All git.

Presentation de l’axe TAL-IA en septembre 2022 pptx

Proposal of a novel unsupervised loss function that provably converges towards the optimal classifier risk and improves the generalization properties of binary deep learning classificiation models. Research paper on HAL

Galaxy detection with Bayesian Neural Networks. Research paper on HAL

Study of the geometrical properties of the loss landscape of growing neural networks: we show that optima are flatter, and improve generalization performances. Research paper on HAL

OLKi videos

AI Charter