QwQ-Med-3

QwQ-Med-3 is a medical reasoning model fine-tuned from Qwen/QwQ-32B on up to three-hop reasoning paths derived from a medical Knowledge Graph. It is introduced in the paper "Bottom-up Domain-specific Superintelligence: A Reliable Knowledge Graph is What We Need" by Bhishma Dedhia, Yuval Kansal, and Niraj K. Jha.

Paper & Code

• 📄 Paper: arXiv:2507.13966
• 💻 Code: jha-lab/bottom-up-superintelligence
• 🌐 Website: kg-bottom-up-superintelligence.github.io
• 📊 Benchmark: yuvalkansal/KG-Med-Bench

Model Description

QwQ-Med-3 is trained using an SFT pipeline:

1. Supervised Fine-Tuning (SFT): The base QwQ-32B model is fine-tuned on question-answer pairs grounded in multi-hop medical knowledge graph paths, teaching the model to produce structured chain-of-thought reasoning aligned with KG-derived evidence.

Intended Use

• Medical question answering requiring multi-hop reasoning
• Evaluation on ICD-coded clinical vignettes
• Research into knowledge graph-guided language model training

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "yuvalkansal/QwQ-Med-3",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("yuvalkansal/QwQ-Med-3")

prompt = "A 45-year-old patient presents with..."
messages = [{"role": "user", "content": prompt}]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

If you use this model, please cite:

@misc{dedhia2025bottomupsuperintelligence,
  author = "{Dedhia, Bhishma and Kansal, Yuval and Jha, Niraj K.}",
  title = "Bottom-up Domain-specific Superintelligence: A Reliable Knowledge Graph is What We Need",
  year = "2025",
  url = {https://arxiv.org/abs/2507.13966}
  }