Llama3-Preferred-MedSwallow-70B Overview

Llama3-Preferred-MedSwallow-70B is a 70 billion parameter language model developed by Preferred Networks, Inc. It is built upon the tokyotech-llm/Llama-3-Swallow-70B-v0.1 base model, which itself is derived from the Llama 3 architecture. A key differentiator for this model is its specialized continued pretraining on an original corpus of medical-related text, making it highly proficient in the medical domain.

Key Capabilities and Performance

This model demonstrates strong performance in medical knowledge assessment, particularly on the Japanese national medical licensing examinations (IgakuQA). It achieved an average score of 395.2 across exams from 2018 to 2022, outperforming GPT-4 (388.8), the base Llama-3-Swallow-70B-v0.1 (348.6), and Meta-Llama-3-70B (334.6) on this specific benchmark. This indicates its enhanced understanding and reasoning capabilities within a medical context.

Intended Use and Limitations

• Good for: Research purposes involving medical-related text, particularly in Japanese. Its strong performance on medical licensing exams suggests its utility for tasks requiring deep medical knowledge.
• Not intended for: Clinical diagnosis. Users are responsible for ensuring compliance with applicable rules and regulations, as the model is developed for research and not direct clinical application.

This model is released under the META LLAMA 3 COMMUNITY LICENSE.