Qwen3.5-4B-Medical-Reasoning

Qwen3.5-4B-Medical-Reasoning is a fine-tuned 4-billion parameter model optimized for complex clinical reasoning and medical question-answering. By leveraging chain-of-thought (CoT) fine-tuning on medical reasoning traces, the model generates explicit step-by-step diagnostic reasoning inside <think> tags prior to delivering a final medical conclusion.

---

Key Highlights

• Reasoning Capabilities: Fine-tuned using step-by-step medical reasoning datasets to prevent superficial guessing and encourage clinical chain-of-thought.
• Dual Output Modes:
  • Thinking Mode (Default): Produces explicit CoT traces inside <think>...</think> tags for explainable medical reasoning.
  • Direct Mode: Can be forced to bypass reasoning and yield succinct direct answers by pre-filling closed tags (<think>\n</think>).
• Strong Medical Benchmark Accuracy: Outperforms standard base models and many 7B-13B non-reasoning models on board-style medical exams.

---

Evaluation Results

MedQA (USMLE 4-Options Benchmark)

Evaluated on the full 1,273-question test set of GBaker/MedQA-USMLE-4-options using single-pass greedy decoding on an NVIDIA A100 GPU:

Metric	Score / Result
Overall Accuracy	68.34% (870 / 1,273)
Successfully Parsed Answers	97.72% (1,244 / 1,273)
Average Inference Time	5.95 seconds / question
Total Evaluation Time	2.10 hours

Extraction & Response Formatting Breakdown

• Option Text / Pattern Extraction: 96.2%
• Tail Fallback Matching: 1.4%
• Empty / Generation Timeout (max_tokens=1024): 2.0%
• Unparseable / Failed Extractions: 0.3%

---

Training Details

• Base Model: Qwen/Qwen3.5-4B
• Fine-Tuning Dataset: FreedomIntelligence/medical-o1-reasoning-SFT (Subset: -EM, Split: Train)
• Fine-Tuning Framework: Unsloth (LoRA fine-tuning)
• Hardware: 1x NVIDIA A100 (40GB VRAM)
• Prompt Format: Qwen ChatML (<|im_start|>, <|im_end|>)

---

How to Use

1. Thinking Mode (Chain-of-Thought Enabled)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "Kerassy/Qwen3.5-4B-Medical-Reasoning"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto"
)

system_prompt = "You are a clinical expert. Think step-by-step inside <think> tags before providing your final medical answer."
user_prompt = "A 45-year-old male presents with sudden chest pain, diaphoresis, and radiation to the left jaw. What is the most likely diagnosis?"

prompt = (
    f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
    f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
    f"<|im_start|>assistant\n"
)

inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_tokens=1024,
    use_cache=True,
    temperature=0.7,
    top_p=0.9,
    eos_token_id=tokenizer.eos_token_id
)

decoded = tokenizer.batch_decode(outputs, skip_special_tokens=False)[0]
response = decoded.split("<|im_start|>assistant\n")[-1].replace("<|im_end|>", "").strip()

print(response)

2. Direct Mode (Bypassing Reasoning)

To bypass the reasoning phase and generate a direct response, pre-fill the assistant turn with closed empty \n tags:

prompt_direct = (
    f"<|im_start|>system\nYou are a succinct medical assistant.<|im_end|>\n"
    f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
    f"<|im_start|>assistant\n<think>\n</think>\n"
)

inputs_direct = tokenizer([prompt_direct], return_tensors="pt").to("cuda")

outputs_direct = model.generate(
    **inputs_direct,
    max_tokens=1024,
    eos_token_id=tokenizer.eos_token_id
)

decoded_direct = tokenizer.batch_decode(outputs_direct, skip_special_tokens=False)[0]
response_direct = decoded_direct.split("<think>\n</think>\n")[-1].replace("<|im_end|>", "").strip()

print(response_direct)

Intended Use & Limitations

Intended Use

• Medical research and evaluation of chain-of-thought capabilities in compact LLMs.
• AI-assisted clinical reasoning benchmark comparison and analysis.

Medical Disclaimer

IMPORTANT: This model is built for research and evaluation purposes only. It is not a certified medical device and should never be used for direct patient diagnosis, treatment advice, or real-world clinical decision-making. Always consult a qualified medical professional for health-related decisions.