TextSquare: Scaling up Text-Centric Visual Instruction Tuning This paper introduces TextSquare, a text-centric visual instruction tuning model that significantly outperforms existing open-source models and even matches state-of-the-art closed-source models on various benchmarks. The model is trained on Square-10M, a massive, high-quality text-centric visual question answering (VQA) instruction tuning dataset generated using closed-source multimodal large language models (MLLMs). The dataset is constructed through a four-step process: Self-Questioning, Answering, Reasoning, and Evaluation. The Square-10M dataset consists of 3.8 million text-rich images collected from diverse sources, and after data generation and filtering, it includes 9.1 million question-answer pairs and reasoning contexts. The dataset is used to train TextSquare, which achieves superior performance on multiple benchmarks, including OCRBench, where it scores 62.2%. TextSquare also outperforms top-tier models like GPT4V and Gemini in six out of ten text-centric benchmarks. The paper demonstrates that reasoning data is crucial for improving model performance and reducing hallucinations in text-centric VQA tasks. It also reveals the relationship between instruction tuning data scale, convergence loss, and model performance, showing that larger, higher-quality datasets lead to better model performance. TextSquare is built using a vision encoder modified from OpenAI CLIP ViT-L-14-336, an LLM based on InternLM-2, and a projector that semantically aligns vision and text tokens. The model is trained using supervised fine-tuning with Square-10M, achieving exceptional performance on text-centric VQA tasks with 8.6B parameters and an image resolution of 700. The paper also discusses the limitations of the approach, including the high computational cost of training large-scale datasets and the fact that synthetic data still cannot reach human-level performance. Overall, the study provides a data-centric perspective on the role of instruction-tuning data in text-centric VQA, confirming that both the quantity and quality of data are crucial to model performance.TextSquare: Scaling up Text-Centric Visual Instruction Tuning This paper introduces TextSquare, a text-centric visual instruction tuning model that significantly outperforms existing open-source models and even matches state-of-the-art closed-source models on various benchmarks. The model is trained on Square-10M, a massive, high-quality text-centric visual question answering (VQA) instruction tuning dataset generated using closed-source multimodal large language models (MLLMs). The dataset is constructed through a four-step process: Self-Questioning, Answering, Reasoning, and Evaluation. The Square-10M dataset consists of 3.8 million text-rich images collected from diverse sources, and after data generation and filtering, it includes 9.1 million question-answer pairs and reasoning contexts. The dataset is used to train TextSquare, which achieves superior performance on multiple benchmarks, including OCRBench, where it scores 62.2%. TextSquare also outperforms top-tier models like GPT4V and Gemini in six out of ten text-centric benchmarks. The paper demonstrates that reasoning data is crucial for improving model performance and reducing hallucinations in text-centric VQA tasks. It also reveals the relationship between instruction tuning data scale, convergence loss, and model performance, showing that larger, higher-quality datasets lead to better model performance. TextSquare is built using a vision encoder modified from OpenAI CLIP ViT-L-14-336, an LLM based on InternLM-2, and a projector that semantically aligns vision and text tokens. The model is trained using supervised fine-tuning with Square-10M, achieving exceptional performance on text-centric VQA tasks with 8.6B parameters and an image resolution of 700. The paper also discusses the limitations of the approach, including the high computational cost of training large-scale datasets and the fact that synthetic data still cannot reach human-level performance. Overall, the study provides a data-centric perspective on the role of instruction-tuning data in text-centric VQA, confirming that both the quantity and quality of data are crucial to model performance.