winninghealth/WiNGPT2-Llama-3-8B-Chat

WiNGPT2-Llama-3-8B-Chat: Medical Domain LLM

WiNGPT2-Llama-3-8B-Chat is an 8 billion parameter language model developed by winninghealth, built upon the Llama 3 architecture. It is specifically fine-tuned for the medical vertical, aiming to provide intelligent medical Q&A, diagnostic support, and medical knowledge services. The model features an 8192-token context length and has undergone significant Chinese language enhancement and multilingual training.

Key Capabilities & Features

• Medical Domain Expertise: Integrates professional medical knowledge and information for specialized applications.
• Enhanced Chinese Language: Demonstrates strong performance in Chinese medical contexts, as evidenced by WiNEval benchmarks.
• Instruction-Tuned: Fine-tuned with approximately 500,000 instruction-following data points for chat and Q&A.
• Multilingual Support: Designed to handle multiple languages, with a focus on Chinese.
• Custom Prompt Format: Utilizes a specific chat template for multi-turn conversations and system instructions.

Performance Highlights (WiNEval)

• MCKQuiz (Objective Medical Questions): Achieved 65.2% accuracy, significantly outperforming Meta-Llama-3-8B-Instruct (49.8%) in Chinese medical objective questions.
• MSceQA (Subjective Medical Questions): Scored 79.8%, comparable to Meta-Llama-3-70B-Instruct-AWQ (78.6%) on subjective medical questions.

Use Cases

• Medical Q&A: Answering general and specific medical questions.
• Diagnostic Support: Providing information and suggestions related to patient conditions (for reference only).
• Medical Knowledge Retrieval: Accessing and synthesizing medical information.
• Medical Translation: Demonstrated capability for English-Chinese translation in the medical context.

Limitations

As a specialized medical LLM, WiNGPT2 provides information and suggestions for reference only and should not replace professional medical advice, diagnosis, or treatment. Users are advised to consult medical professionals and independently evaluate the information provided.