DeepSeek-R1-Distill-Qwen-7B / KG-Curriculum-SFT
Model Description
Fine-tuned from DeepSeek-R1-Distill-Qwen-7B using LoRA on a
116,859-item KG-grounded multi-hop clinical reasoning curriculum
derived from the UMLS 2025AA Metathesaurus. Developed as part of
the undergraduate thesis (Princeton ECE, 2026).
Training Data
jha-lab/med-si-dataset
116,859 USMLE-style QA pairs with chain-of-thought traces grounded in explicit KG relational paths
(1-hop: 66,579; 2-hop: 50,280). Generated via Gemini 2.5 Flash Lite and validated by DeepSeek-V3.
Training Procedure
- Base model: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
- Method: LoRA (r=16, α=16, dropout=0.05)
- Learning rate: 2e-6, cosine schedule
- Epochs: 4
- Hardware: 2× A100 80GB
Intended Use
Research only. Clinical domain multi-hop reasoning evaluation.
Not validated for clinical deployment.
