OCRBench v2 Benchmark for Visual Text Localization and Reasoning

OCRBench v2 is a benchmark dataset for evaluating large multimodal models, containing at least 10,000 test samples. It includes subsets for English and Chinese, with the English subset containing 7,400 samples. The dataset focuses on visual text localization and reasoning tasks.

Use Cases

• Benchmark large multimodal models on visual text localization tasks using the provided test samples.
• Evaluate model performance on English and Chinese visual text reasoning using the specified EN and CN subsets.
• Assess multimodal model capabilities on tasks defined by the OCRBench v2 benchmark framework.

Strengths

• Contains at least 10,000 test samples for model evaluation.
• Includes a dedicated English subset of 7,400 samples.
• Specifically designed for benchmarking visual text localization and reasoning, a defined task.

Limitations

• The total dataset size, column structure, and sample data details are unknown.
• Limited to benchmark evaluation; may not be suitable for general model training without additional data.
• Geographic and temporal coverage of the source images is unspecified.

Provenance

Source

morpheushoc on Hugging Face, associated with a GitHub repository and arXiv paper.

Collection Method

Curated as a benchmark dataset for evaluating large multimodal models.

Freshness

Last updated on January 29, 2025.