EPFLiGHT/Apertus-70B-MeditronFO
EPFLiGHT/Apertus-70B-MeditronFO is a 70 billion parameter medical specialist LLM developed by EPFLiGHT, specializing Apertus-70B-Instruct on the Fully Open Meditron Corpus. This model is part of the Fully Open Meditron family, offering an end-to-end auditable pipeline for clinical LLMs with open weights, data, and training. It establishes a new state-of-the-art in medical accuracy among fully open medical LLMs, outperforming its base model on standard medical benchmarks. The model is primarily intended for medicine-related tasks and evaluation, serving as an assistive tool in clinical contexts.
Loading preview...
Apertus-70B-MeditronFO: A Fully Open Medical LLM
Apertus-70B-MeditronFO is EPFLiGHT's flagship medical specialist Large Language Model, built upon the Apertus-70B-Instruct base model. It is a key component of the Fully Open Meditron initiative, which emphasizes an end-to-end auditable pipeline for clinical LLMs, featuring open weights, data, and training methodologies.
Key Capabilities and Differentiators
- Medical Specialization: Fine-tuned on the extensive Fully Open Meditron Corpus, comprising 601k examples (approximately 150M tokens) from aggregated public medical QA datasets and clinician-vetted synthetic components.
- State-of-the-Art Medical Accuracy: Achieves new benchmarks in medical accuracy among fully open medical LLMs, demonstrating significant improvements across standard medical benchmarks like MedMCQA (+3.89%), MedQA (+7.94%), PubMedQA (+8.40%), and HealthBench Hard (+7.86%).
- Clinician-Validated Performance: Preferred over its base model, Apertus-70B, in 85.3% of comparisons based on the clinician-validated LM judge evaluation, AutoMOOVE.
- Open and Auditable: Provides full transparency with open weights, open data, and an open training recipe, facilitating scrutiny and trust in clinical applications.
Intended Use and Limitations
This model is specifically designed for medicine-related tasks and evaluation, serving as an assistive tool. Users should be aware that while specialized, the generated content may not always be factually accurate or free from biases, necessitating critical evaluation and verification of important information.