zwyang6/

Perception_ZwZ-RL-VQA

This dataset contains 74K high-quality VQA pairs generated via Region-to-Image Distillation (R2I) for training multimodal large language models (MLLMs) on fine-grained perception tasks without test-time tool use.

The Zooming without Zooming (ZwZ) method transforms "zooming" from an inference-time tool into a training-time primitive:

Zoom-in Synthesis: Strong teacher models… See the full description on the dataset page: https://huggingface.co/datasets/zwyang6/Perception_ZwZ-RL-VQA.