JUDO (ICLR 2026)

Official implementation of
"JUDO: A Juxtaposed Domain-oriented Multimodal Reasoner for Industrial Anomaly QA"

Base model: Qwen2.5-VL-7B
Official trained checkpoint:
https://huggingface.co/woodavid31/JUDO

conda create -n judo python=3.10
conda activate judo
bash setup.sh

Multi-GPU is recommended for GRPO training.

The training script performs:

• Segmentation SFT
• Domain knowledge SFT
• GRPO alignment

cd open-r1-multimodal
bash seg_sft_grpo.sh

Download datasets:

MMAD
https://huggingface.co/datasets/jiang-cc/MMAD

REAL-IAD
https://huggingface.co/datasets/Real-IAD/Real-IAD/tree/main/realiad_512

Place datasets according to the paths expected in the training scripts.

cd eval
python eval_seg_mult.py

To evaluate the official JUDO model, set the model path to:

./JUDO

Or replace it with your locally trained checkpoint.

Model outputs follow:

<seg>...</seg>
<think>...</think>
<answer>...</answer>

This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2025-02653113, High-Performance Research AI Computing Infrastructure Support at the 2 PFLOPS Scale)