YDXX/G-Health-14B-instruct

Overview

G-Health-14B-instruct is a specialized large language model from the G-Health family, built upon the Qwen3 architecture and fine-tuned for medical and preventive health applications. This model is particularly adept at interpreting health checkup reports, providing actionable insights and personalized recommendations.

Key Capabilities

• Medical Domain Alignment: Undergoes two-stage alignment (SFT and DPO) using millions of medical dialogue and preference samples to enhance robustness and communication quality in medical contexts.
• Health Checkup Report Specialization: Further fine-tuned on health checkup report data to improve:
  • Interpretation of lab values and imaging conclusions.
  • Cautious risk signaling, even under uncertainty.
  • Enhanced personalization for tailoring explanations and recommendations.

Training Highlights

Starting from Qwen3, the base models (G-Health-14B-Base) are aligned using 2.8 million supervised fine-tuning (SFT) dialogue samples and 1.6 million direct preference optimization (DPO) samples. The instruct version is then built on these base models with additional fine-tuning specifically for health checkup report data.

Good For

• Generating structured interpretations from health checkup reports.
• Providing personalized health recommendations based on individual data.
• Applications requiring robust and high-quality communication in medical and preventive health scenarios.