configs:
- config_name: default
data_files:
- split: sft
path: sft/*.parquet
- split: rl
path: rl/*.parquet
- split: images
path: images/*.parquet
UniReason-Med Data
Training data release for the UniReason-Med medical reasoning artifact.
Overview
This dataset contains SFT and GRPO/RL training mixtures for a medical multimodal reasoning framework for 2D and 3D medical image understanding. The examples train models to perform interleaved image-text reasoning with region grounding, where the model can identify image regions with bounding-box tokens, inspect the corresponding crop, and continue reasoning before producing the final answer.
Image Release
The 2D image data is included in the images split as Parquet records with image bytes and neutral relative paths. SFT and RL records refer to these images through image_ids. The 3D samples are released as text-only records. We do not redistribute 3D image data because these samples are derived from M3D, whose underlying image sources include Radiopaedia and may require separate permission. Users who need the 3D images should obtain the necessary authorization from the original data providers, including M3D/Radiopaedia where applicable.