researcher2026/OpenMedReason

This data is the Open-PMC derived data part that we discuss in the paper

Medical VQA with Reasoning Traces (Anonymous)

A multiple-choice medical visual question answering benchmark with chain-of-thought reasoning traces. Each example consists of a medical image (radiology, pathology, clinical photograph, etc.), a multiple-choice question with labeled options, a reasoning trace, and the correct answer letter.

This dataset is released anonymously in support of a peer-reviewed submission.

Columns

Column	Type	Populated on	Description
image	Image	train, test	The medical image relevant to the question (JPEG, embedded).
question	string	train, test	Question text followed by labeled answer options (e.g. "A. ...").
reasoning	string	train, test	Chain-of-thought reasoning trace that supports the correct answer.
answer	string	train, test	Single uppercase letter giving the correct option (e.g. "A").
Perception	string	test only	JSON list of observation-axis unit-questions (what the image actually shows).
Medical knowledge	string	test only	JSON list of knowledge-axis unit-questions (relevant clinical/medical facts).
Rationale	string	test only	JSON list of inference-axis unit-questions (reasoning-from-observations claims).

Perception / Medical knowledge / Rationale are evaluation aids: each is a JSON-encoded list of {topic, claim, presence_question, correctness_question, source_quote, importance} items derived from the reference reasoning trace. They are populated for the test split only and held as "[]" strings on the train split for schema parity (so the dataset viewer works on both splits).

Loading

from datasets import load_dataset
import json

ds = load_dataset("researcher2026/OpenMedReason")
print(ds)
print(ds["train"][0])

# Eval-time helpers (test split):
ex = ds["test"][0]
perception = json.loads(ex["Perception"])
knowledge  = json.loads(ex["Medical knowledge"])
rationale  = json.loads(ex["Rationale"])
print(f"perception items: {len(perception)}")

Anonymity

The dataset is hosted from an anonymous account and the dataset card omits any identifying information about the authors, institution, or project.