Qwen2.5-0.5B-MedReason-SFT

A 494M parameter language model fine-tuned for structured clinical reasoning and medical question answering. Built on Qwen-2.5-0.5B-Instruct using QLoRA and Chain-of-Thought (CoT) supervision, this model demonstrates that compact models can produce structured, step-by-step medical diagnostic reasoning when trained on high-quality clinical data.

---

Demo

The Gradio interface for submitting clinical queries and reviewing AI-generated diagnostic reasoning output.

---

Model Details

Summary

Property	Value
Model Name	Qwen2.5-0.5B-MedReason-SFT
Base Model	Qwen/Qwen2.5-0.5B-Instruct
Parameters	494 Million
Architecture	Transformer-based Causal Decoder
Training Method	Supervised Fine-Tuning (SFT) with QLoRA
Quantization (Training)	4-bit NormalFloat (NF4)
Merged Precision	16-bit (FP16)
Context Window	2048 Tokens
License	Apache 2.0

Model Description

This model was fine-tuned to internalize structured medical reasoning through Chain-of-Thought (CoT) data derived from the Baichuan-M3-235B model. The training data includes a reasoning_content field that teaches the model to think through differential diagnoses, symptom analysis, and clinical decision pathways before producing a final answer.

The project demonstrates that sub-1B parameter models can be meaningfully specialized for clinical domains without large compute budgets, using efficient fine-tuning techniques (LoRA, 4-bit quantization, gradient accumulation) on a single consumer-grade GPU.

---

Training Details

Dataset

Property	Value
Dataset	OpenMed Medical-Reasoning-SFT
Source	Derived from Baichuan-M3-235B
Training Subset Used	124,520 samples
Full Dataset Size	~1,790,000 samples
Format	Conversational JSONL with reasoning_content CoT field
HuggingFace Link	openmed/Medical-Reasoning-SFT

The dataset contains medical question-answer pairs enriched with reasoning chains that simulate expert-level clinical thinking. Each sample follows an instruction-response structure, with the reasoning trace embedded to guide the model toward interpretable, step-by-step outputs.

Hardware

• GPU: NVIDIA Tesla T4 (single GPU)
• Training Library: Unsloth
• Inference Speed Gain: ~2x via Unsloth Fast Inference kernels

Hyperparameters

Parameter	Value
Learning Rate	2e-4
Optimizer	AdamW (8-bit)
Batch Size (per device)	1
Gradient Accumulation Steps	4
Effective Batch Size	4
LoRA Rank (R)	16
LoRA Alpha	16
Weight Decay	0.01
Quantization	NF4 (4-bit)

Training Metrics

Metric	Value
Initial Training Loss	2.3664
Final Training Loss (1,000 steps)	1.6457
Loss Reduction	~30.5%

---

How to Use

Installation

pip install torch torchvision torchaudio
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
pip install transformers

Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Rumiii/Qwen2.5-0.5B-MedReason-SFT"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {
        "role": "user",
        "content": "A 58-year-old male presents with crushing chest pain radiating to the left arm, diaphoresis, and shortness of breath. What is the most likely diagnosis and immediate management?"
    }
]

input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)

response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(response)

Gradio Interface

A dedicated Gradio agent is included in the source repository for interactive clinical review:

git clone https://github.com/sufirumii/Medical-Reasoning-AI-Agent-Fine-Tuning-Qwen-2.5-0.5B
cd Medical-Reasoning-AI-Agent-Fine-Tuning-Qwen-2.5-0.5B
python gradio_agent.py

---

Intended Use

Appropriate Use Cases

• Medical education and student training aids
• Clinical reasoning research and benchmarking
• Exploring the capabilities of compact fine-tuned models in healthcare NLP
• Prototyping AI-assisted diagnostic tools for research environments

Out-of-Scope Use Cases

• Direct clinical decision-making without physician oversight
• Replacing licensed medical professionals
• Deployment in any production healthcare system without rigorous validation
• Use in emergency or life-critical medical situations

---

Limitations

• Scale: Trained on an initial subset (124,520 samples) of the full 1.79M sample dataset. Performance may improve significantly with full-dataset training.
• Parameter Count: At 494M parameters, the model may lack the depth required for nuanced or rare clinical presentations that larger models handle more reliably.
• Hallucination Risk: Like all language models, this model can produce confident but incorrect medical statements. All outputs must be validated by a qualified clinician.
• Language: Trained on English-language data only. Performance in other languages is not guaranteed.
• Recency: Medical knowledge has a training cutoff and does not reflect the latest clinical guidelines or drug approvals.

---

Bias and Ethical Considerations

The training data is derived from a large-scale model (Baichuan-M3-235B) and may reflect biases present in that model or its underlying sources. Medical AI systems are known to exhibit demographic bias — including disparities across age, sex, ethnicity, and socioeconomic status — which may affect the quality of reasoning for underrepresented patient populations. Users should treat all outputs critically and not apply them uniformly across diverse patient groups without independent clinical assessment.

---

Clinical Disclaimer

This model is intended strictly for research and educational purposes. It is not approved for clinical use and must not be used as a substitute for professional medical advice, diagnosis, or treatment. All AI-generated medical reasoning must be reviewed and verified by a licensed and qualified healthcare professional before any clinical consideration.

---

Citation

If you use this model in your research, please cite it as:

@misc{qwen25_medreason_sft,
  title        = {Qwen2.5-0.5B-MedReason-SFT: A Compact Model for Clinical Chain-of-Thought Reasoning},
  author       = {Rumi Sufi},
  year         = {2026},
  howpublished = {HuggingFace},
  url          = {https://huggingface.co/Rumiii/Qwen2.5-0.5B-MedReason-SFT}
}

---

Acknowledgements

• Unsloth for memory-efficient fine-tuning infrastructure
• OpenMed for the Medical-Reasoning-SFT dataset
• Qwen Team / Alibaba Cloud for the Qwen-2.5 base model family