BioMistral: Open-Source Pretrained LLMs for the Medical Domain (arXiv 2402.10373)

Forum topic · 小凯 · 2026-07-05

Summary

BioMistral is a collection of fully open-source large language models specialized for the biomedical and medical domain, presented by a team from the University of Nantes, LIRICA/LS2N laboratories, and collaborators in February 2024 on arXiv (2402.10373). The suite is built by continued pretraining of Mistral 7B on the complete English corpus of PubMed Central, yielding BioMistral 7B. Beyond the English base model, the authors release multilingual medical variants covering eight additional languages, produced by further adapting BioMistral to PubMed Central excerpts in each target language. The models are quantized for lightweight, locally deployable use, and an instruction-tuned variant is created by fine-tuning on Alpaca biomedical synthetic instruction data. Evaluation spans ten established medical benchmarks, where BioMistral outperforms other open-source 7B-scale medical models and approaches larger closed-source counterparts, while also showing that chain-of-thought reasoning remains challenging for small medical LLMs. All models, data, code, and evaluation pipelines are released publicly, making BioMistral a reference toolkit for transparent, reproducible research and privacy-preserving clinical AI applications.

BioMistral: Open-Source Pretrained LLMs for the Medical Domain

> Source: https://arxiv.org/abs/2402.10373 (February 2024) > Authors: Yanis Labrak, Adrien Bazoge, Emmanuel Morin, Pierre-Antoine Gourraud, Mickael Rouvier, Richard Dufour

Overview

BioMistral is a suite of fully open-source large language models tailored to the biomedical domain. It is built by continued pretraining of Mistral 7B on the complete English PubMed Central corpus, producing the base model BioMistral 7B.

Key contributions

BioMistral 7B: continued pretraining of Mistral 7B on all English PubMed Central abstracts and articles.
Multilingual variants: additional models adapted to PubMed Central excerpts in eight other languages, enabling medical NLP beyond English.
Quantization: quantized versions for lightweight, local deployment on commodity hardware.
Instruction-tuned model: created by fine-tuning BioMistral on Alpaca biomedical synthetic instruction data.
Open release: all models, training data, code, and evaluation pipelines are public.

Evaluation

BioMistral was evaluated on 10 established medical benchmarks. Key findings:

BioMistral outperforms other open-source 7B-scale medical models.
It approaches the performance of larger closed-source medical LLMs on several tasks.
Chain-of-thought (CoT) reasoning remains difficult for small medical LLMs: prompt-level CoT strategies did not consistently improve results and can even hurt performance.

Why it matters

Fully open medical LLMs support privacy-preserving, locally deployable clinical applications and transparent, reproducible research — addressing the licensing and data-governance limits of closed models in healthcare.

Limitations

7B scale limits complex reasoning abilities.
English-centric continued pretraining for the base model; multilingual quality varies by language.

Related entries

Better to Ask in English: Cross-Lingual Evaluation of Large Language Models

References

Original paper: BioMistral: A Collection of Open-Source Pretrained Large Language Models for Medical Domains. arXiv:2402.10373