AI Builds|Live
Med360 Research Paper
Research paper on fine-tuning LLMs for Indian healthcare - introducing the first Hinglish medical conversation dataset.
Overview
Med360 research paper presents a family of fine-tuned large language models specifically designed for Indian healthcare contexts. The paper addresses critical gaps in existing medical AI systems including lack of Indian medical context, absence of Hinglish support, and unfamiliarity with Indian pharmaceutical names. Key insight: automated metrics may penalize clinically useful concise responses, highlighting the need for domain-specific evaluation methods.
Specs
Model FamilyMed360 Lite (4B), Pro (12B), Ultra (27B)
Training Corpus173,000+ curated medical examples
MethodologyMulti-stage fine-tuning with LoRA
Key ContributionFirst synthetic Hinglish medical conversation dataset
Features
- Multi-stage fine-tuning approach using LoRA
- Stage 1: Core conversational abilities (23,400 examples)
- Stage 2: Expanded medical knowledge (95,729 examples)
- Stage 3: Indian healthcare specialization (54,007 examples)
- First synthetic Hinglish medical conversation dataset
- Superior clinical utility for Indian healthcare contexts
- Handles Indian pharmaceutical names and disease protocols