Ask or search anything...

Researchers at BAAI and collaborating institutions developed Infinity-MM, an open-source multimodal instruction dataset comprising tens of millions of high-quality samples, alongside a scalable data synthesis method utilizing open-source VLMs. Training the 2-billion-parameter Aquila-VL-2B model on Infinity-MM resulted in state-of-the-art performance among similarly scaled models across numerous visual question answering, knowledge, and mathematical reasoning benchmarks.