Cerebras Systems, an AI computer maker, has partnered with M42, an operator of healthcare facilities, to develop an open-source large language model (LLM) called Med42 for health applications. Med42 is a refinement of Llama 2, an open-source LLM released by Meta Properties, and it has been fine-tuned using a health-related dataset compiled by the companies. The program aims to assist healthcare providers with tasks such as medical question answering, patient record summarization, medical diagnosis, and general health Q&A. The Med42 neural net has been fine-tuned with a dataset of 700,000 question-and-answer pairs from publicly available sources, and it does not include patient data or personally identifiable information.
The Med42 code is now available on HuggingFace, along with performance data. The companies plan to release further enhancements as they collaborate with healthcare professionals to improve the model’s capability and performance. The fine-tuning of Med42 was done on Condor Galaxy, an AI computer built by Cerebras, which is considered the world’s largest supercomputer for AI. The vast memory and AI cores in Condor Galaxy enabled rapid setup and reduced training time for Med42.
In terms of performance, Med42 achieves competitive results on various medical benchmarks, surpassing the prior state of the art among openly available medical LLMs. On the US Medical Licensing Examination sample exam, Med42 achieves a 72% accuracy, outperforming OpenAI’s closed-source GPT 3.5. However, it falls short of GPT4’s 84.3% accuracy.
Cerebras has been actively involved in open-source projects, having previously published several versions of generative AI programs for unrestricted use. In August, the company released Jais-Chat, the world’s most powerful Arabic-language LLM, as an open-source program.
While Med42 is not currently in production, the companies plan to make it available for clinical deployment following successful testing. They prioritize the safe application of the technology and commit to extensive safety evaluation before rolling it out. Med42 can be deployed on-premise, customized to healthcare providers’ needs, and designed to limit external intrusions.
Overall, Cerebras and M42’s collaboration on Med42 demonstrates an innovative approach to leveraging AI in healthcare, offering potential benefits for healthcare providers in various applications.