Preferred-MedLLM-Qwen-72B: A Specialized Medical LLM

Preferred-MedLLM-Qwen-72B is a 72.7 billion parameter language model developed by Preferred Networks, Inc. It is a fine-tuned version of the Qwen/Qwen2.5-72B architecture, specifically enhanced through continued pretraining on a unique corpus of medical-related text.

Key Capabilities & Performance

This model demonstrates exceptional proficiency in medical knowledge, particularly within the context of Japanese medical examinations. It has been rigorously evaluated on the Japanese medical licensing examination (IgakuQA) from 2018 to 2022, where it achieved an average score of 431.2. This performance surpasses several leading models, including GPT-4o (430.4), its base model Qwen2.5-72B (398.4), and GPT-4 (388.8).

Training & Architecture

The model leverages the robust Qwen2.5-72B architecture and benefits from a specialized continued pretraining phase using an original medical text dataset. This targeted training allows it to excel in medical question-answering and reasoning tasks. It supports a substantial context length of 131072 tokens.

Limitations

It is important to note that Preferred-MedLLM-Qwen-72B was developed for research purposes only and is not intended for clinical diagnosis. Users are responsible for ensuring compliance with all applicable rules and regulations when deploying this model.