justinj92/MediQwen-Reasoning-4B

• Developed by: justinj92
• License: apache-2.0
• Finetuned from model : unsloth/Qwen3-4B-Instruct-2507
• GPU : AMD MI300x
• EPOCH : 2
• Training Time : 3 Days

Dataset Mix

• Reasoning - II-Medical-Reasoning at 70%
• Non-Reasoning - Mediflow at 30%

Benchmark	MediQwen-Reasoning-4B	Qwen3-4B-Instruct-2507	Δ
MedQA	55.85% (711/1273)	12.73% (162/1273)	+43.12%
MedMCQA	54.29% (2271/4183)	53.50% (2238/4183)	+0.79%
PubMedQA	71.90% (719/1000)	60.00% (600/1000)	+11.90%

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.