sczzz/RaDaR-32B

RaDaR-32B: Rare Disease Differential Diagnosis LLM

RaDaR (Rare Disease navigatoR) is a 32.8 billion parameter large language model developed by sczzz, specifically engineered for differential diagnosis of rare diseases. It is initialized from DeepSeek-R1-Distill-Qwen-32B and has a context length of 32768 tokens.

Key Capabilities

• Direct Free-Text Processing: RaDaR processes raw clinical narratives, such as patient history, symptoms, lab results, and imaging findings, without requiring prior conversion to structured formats like Human Phenotype Ontology (HPO) terms.
• Specialized Reasoning: The model is fine-tuned on a corpus of 49,170 real-world and 104,666 phenotype-anchored synthetic rare-disease cases, using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to enhance diagnostic reasoning.
• Local Deployment Focus: Designed for institutional use, RaDaR supports local deployment to keep sensitive clinical data within an organization's computing environment, addressing data privacy concerns.
• Differential Diagnosis: It generates a ranked list of potential rare-disease diagnoses, intended as a clinical decision-support tool to prioritize diseases for clinician review, rather than providing definitive diagnoses.

Use Cases

• Clinical Decision Support: Assisting healthcare professionals in identifying potential rare diseases from complex patient records.
• Research: Facilitating research into rare disease diagnostics by providing a specialized reasoning model.
• Secure Environments: Ideal for settings where clinical data privacy and local processing are paramount, such as hospitals and research institutions.