NorthernTribe-Medical-Reasoning-7B

Model Description

NorthernTribe-Medical-Reasoning-7B is a state-of-the-art medical reasoning model fine-tuned on the Comprehensive HealthBench v2 dataset.

It is built upon Qwen-2.5-7B-Instruct, selected for its superior performance in mathematics and reasoning benchmarks compared to Llama 3. The model is specifically optimized to provide Chain-of-Thought (CoT) reasoning for complex medical inquiries, differential diagnoses, and clinical decision support.

Key Features

• Medical Chain-of-Thought: Trained to "think" before answering, providing detailed rationale for every medical decision.
• Massive Knowledge Base: Fine-tuned on 1.4 million high-quality medical instruction-response pairs, including SOTA datasets like ReasonMed, Medical-o1, and PubMedQA.
• Instruction Following: Capable of handling complex multi-turn clinical scenarios.

Intended Use

This model is designed for:

• Clinical Decision Support: Assisting healthcare professionals with differential diagnosis and treatment planning.
• Medical Education: Explaining complex physiological mechanisms and disease pathologies.
• Biomedical Research: Summarizing and reasoning over medical literature.

Disclaimer: This model is for research and educational purposes only. It is not a substitute for professional medical advice, diagnosis, or treatment. Always consult with a qualified healthcare provider for clinical decision-making.

Training Details

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Dataset: NorthernTribe-Research/comprehensive-healthbench-v2 (1.4M examples)
• Method: QLoRA Fine-tuning via Unsloth
• SOTA Techniques:
  • NEFTune (Noisy Embeddings Fine Tuning): Implemented to improve model generalization and robustness.
  • Teacher-Student Distillation: System prompts strictly enforce the model to act as an "Expert Medical Teacher", engaging in deep Chain-of-Thought reasoning for every response.
• Infrastructure: Trained on Nvidia A100/H100 GPUs.

Usage

Inference (Python)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "NorthernTribe-Research/NorthernTribe-Medical-Reasoning-7B",
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

prompt = """<|im_start|>system
You are a highly capable medical AI assistant. Answer the following question accurately and provide detailed reasoning where appropriate.<|im_end|>
<|im_start|>user
A 45-year-old male presents with chest pain... [Clinical Vignette]... What is the most likely diagnosis?<|im_end|>
<|im_start|>assistant
"""

inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens = 512, use_cache = True)
print(tokenizer.batch_decode(outputs)[0])

Citation

If you use this model, please cite:

@misc{northerntribe2025medical,
  title={NorthernTribe-Medical-Reasoning-7B: A Chain-of-Thought Medical LLM},
  author={NorthernTribe Research},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/NorthernTribe-Research/NorthernTribe-Medical-Reasoning-7B}}
}