EPFLiGHT/Apertus-8B-MeditronFO
EPFLiGHT/Apertus-8B-MeditronFO is an 8 billion parameter medical specialist large language model developed by LiGHT, fine-tuned from Apertus-8B-Instruct. It is part of the Fully Open Meditron family, offering an end-to-end auditable pipeline for clinical LLMs with open weights, data, and training. This model achieves 48.26% average accuracy across five medical benchmarks, outperforming its base model and other small fully open medical LLMs, making it ideal for medical question answering and clinical assistance.
Loading preview...
Apertus-8B-MeditronFO: A Fully Open Medical LLM
Apertus-8B-MeditronFO is an 8 billion parameter medical specialist LLM developed by EPFL's LiGHT laboratory. It is a medical specialization of the Apertus-8B-Instruct model, fine-tuned on the extensive Fully Open Meditron Corpus. This model is a key component of the Fully Open Meditron initiative, which aims to provide the first end-to-end auditable pipeline for clinical LLMs, ensuring transparency with open weights, data, training recipes, and clinician-vetted corpus construction.
Key Capabilities and Performance
- Leading Medical Accuracy: Achieves 48.26% average accuracy across five medical benchmarks (MedMCQA, MedQA, PubMedQA, MedXpertQA, HealthBench), surpassing Apertus-8B-Instruct's 42.39%.
- Clinician-Validated Preference: Preferred over its base model in 74.2% of comparisons on AutoMOOVE, an LLM-judge evaluation using 516 clinician-written vignettes.
- Significant HealthBench Improvement: Demonstrates the largest HealthBench gain within the MeditronFO family, improving by +16.1 points (from 26.0 to 42.1).
- Comprehensive Training Data: Trained on 533,888 examples from the Fully Open Meditron Corpus, including curated QA, synthetic QA, guidelines QA, and synthetic vignettes, with answers generated by gpt-oss-120b and clinician co-authored prompts.
Intended Use
Apertus-8B-MeditronFO is specialized for medicine-related tasks and evaluations. It is designed as an assistive tool for medical professionals and researchers, providing support for differential diagnosis, medical question answering, and other clinical applications. Users should always verify generated information and critically evaluate content, as the model is not a definitive source.