DatasetMan/

PM4Bench-QGO-Train

Synthetic multilingual OCR data for QGO reinforcement learning

The paper Benchmarking and Boosting Multilingual Capabilities of LVLMs via OCR-Centric Reinforcement Learning uses PM4Bench to show that OCR is a key source of cross-lingual performance gaps when text is rendered visually. QGO addresses that finding with GRPO on synthetic OCR data, without task-specific VQA or GUI supervision. This repository contains the… See the full description on the dataset page: https://huggingface.co/datasets/DatasetMan/PM4Bench-QGO-Train.