Etienne Boisson
Research Associate at EPFL. Formerly at Yale University, where I contributed to research on LLM training and data generation. I work at EPFL’s LiGHT lab, focused on using technology to save lives. I’m driven by the belief that AI can expand global access to knowledge.
For the past two years I’ve been part of Meditron, an academic initiative creating SOTA medical LLMs. My contributions include:
- Meditron Expansion, expanding the Medical Meditron protocol to different family of models such as Gemma, Phi, Llama and Qwen.
- Llama-3-Meditron, an Open-Weight Suite of Medical LLMs Based on Llama-3.1
News
Experience
[2025-present] Research Associate. Working on the new iteration of the Meditron protocol.
[2024-2025] Research Associate. Co-led the Meditron conversation team and expanding the Meditron protocol.
[2023-2025] Research Assistant. Working on Llama-3-Meditron.
[2022-2025] MSc in Computer Science & Data Science at EPFL, GPA of 5,5/6.
[summer 2024] AI Intern Engineer at the International Commitee of the Red Cross.
[summer 2023] AI Intern Enginee at Synature.
[2018-2023] BSc in Computer Science at EPFL.
Publications / Projects
Expanding the Meditron Protocol – Cross-Family Adaptation, Safety and Conversational Capabilities in Generative Medical Decision Support Systems
E. Boisson, M. Jaggi, M.-A. Hartley.
Master thesis
Pdf
We enhance the safety, ethics, instruction-following and conversation capabilities of Meditron LLMs through fine-tuning, synthetic data training and evaluation and adapt the Meditron Protocol to various foundation models on a complex infrastructure.
E. Boisson, M. Jaggi, M.-A. Hartley.
Master thesis
We enhance the safety, ethics, instruction-following and conversation capabilities of Meditron LLMs through fine-tuning, synthetic data training and evaluation and adapt the Meditron Protocol to various foundation models on a complex infrastructure.
Llama-3-Meditron: An Open-Weight Suite of Medical LLMs
A. Sallinen, A. Solergibert, M. Zhang, G. Boyé, M. Dupont-Roc, X. Theimer-Lienhard, E. Boisson, B. Bernath, H. Hadhri, A. Tran, et al.
AAAI 2025, Gen AI for health
OpenReview
We finetune Llama-3.1 with the Meditron mixture using SFT and ORPO. Meditron outperforms its base model by 3% on medical benchmarks, and our Meditree inference method allows us to reach GPT4-Base performance of 80% accuracy on medical benchmarks, a 5% gain over Meditron alone.
A. Sallinen, A. Solergibert, M. Zhang, G. Boyé, M. Dupont-Roc, X. Theimer-Lienhard, E. Boisson, B. Bernath, H. Hadhri, A. Tran, et al.
AAAI 2025, Gen AI for health
OpenReview
We finetune Llama-3.1 with the Meditron mixture using SFT and ORPO. Meditron outperforms its base model by 3% on medical benchmarks, and our Meditree inference method allows us to reach GPT4-Base performance of 80% accuracy on medical benchmarks, a 5% gain over Meditron alone.
GPoeT: a language model trained for rhyme generation on synthetic data
A. Popescu-Belis, A. R. Atrio, B. Bernath, E. Boisson, T. Ferrari, X. Theimer-Lienhard, G. Vernikos.
Proceedings of the 7th Joint SIGHUM Workshop on Computational Linguistics for CH, SSH and Lit.
ACLAnthology
We finetune GPT-2 on 142 MB of natural poems and 6 MB of rhyming poems and find that we obtain rhymes 60% of the time versus the 11% baseline.
A. Popescu-Belis, A. R. Atrio, B. Bernath, E. Boisson, T. Ferrari, X. Theimer-Lienhard, G. Vernikos.
Proceedings of the 7th Joint SIGHUM Workshop on Computational Linguistics for CH, SSH and Lit.
ACLAnthology
We finetune GPT-2 on 142 MB of natural poems and 6 MB of rhyming poems and find that we obtain rhymes 60% of the time versus the 11% baseline.