Abstract

Vision-Language Models (VLMs) have recently made significant progress, butthe limited scale and quality of open-source instruction data hinder theirperformance compared to closed-source models. In this work, we address thislimitation by introducing Infinity-MM, a large-scale multimodal instructiondataset with 40 million samples, enhanced through rigorous quality filteringand deduplication. We also propose a synthetic instruction generation methodbased on open-source VLMs, using detailed image annotations and diversequestion generation. Using this data, we trained a 2-billion-parameter VLM,Aquila-VL-2B, achieving state-of-the-art (SOTA) performance for models ofsimilar scale. This demonstrates that expanding instruction data and generatingsynthetic data can significantly improve the performance of open-source models.