A small fine-tuned Qwen3 model that takes a Pathology medical multiple-choice question (MCQ) and predicts which topic within Pathology the question is about.

This is a per-subject topic classifier — the second hop in the MedMCQ three-hop pipeline. It assumes the input MCQ has already been routed to Pathology by the subject classifier.

The MedMCQ pipeline

The MedMCQ project explores small, specialized models for medical MCQs. Instead of using one large model for everything, it splits the task across three hops:

Subject routing — the subject classifier picks the medical subject.
Topic classification — this model. Given a Pathology MCQ, pick the topic within Pathology.
Answer generation — the Pathology generator produces the answer / new MCQs in Pathology.

Each hop is a separate, narrow model. They are all published under the MedMCQ Medical Models collection.

Quick start

python

from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "stravoris/medmcq-pathology-classifier-qwen3-0.6b"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)

prompt = """Classify the following Pathology MCQ by topic.

Question: <a Pathology MCQ here>
A) <option A>
B) <option B>
C) <option C>
D) <option D>

Topic:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=16, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Prompt format

The model expects MCQs in this exact form:

Classify the following Pathology MCQ by topic.

Question: <question text>
A) <option A>
B) <option B>
C) <option C>
D) <option D>

Topic:

The model completes the prompt with one of the topic labels seen during training. The full topic list is in the dataset card.

What this model is not

This is a sample model for demonstration. It is not a production-grade medical AI system:

It has not been formally evaluated against board-level benchmarks.
It should not be used to make clinical decisions or provide medical advice.
It is narrow: it only understands Pathology MCQs and assumes the upstream subject router did its job.
It was trained on a curated educational dataset and inherits any biases or gaps in that data.

The MedMCQ project exists to explore small-model pipeline architectures for medical reasoning, not to ship a medical product.

Training data

Trained on the Pathology subset of the Stravoris Medical MCQ dataset — educational Pathology MCQs labeled by topic.

Base model

Fine-tuned from Qwen/Qwen3-0.6B. The base model's license and usage terms also apply.

License

Apache 2.0. See LICENSE.

Collection

Part of the MedMCQ Medical Models collection — all 31 models that make up the MedMCQ three-hop pipeline.