OCR-Quality: 1,000 Document Images with Human-Annotated OCR Scores

1,000 document images with human-annotated quality scores and OCR text extracted by the Qwen2.5-VL-72B model. The dataset was created by Aslan-mingye and last updated on October 17, 2025. It is designed as a benchmark for evaluating OCR performance across diverse document types including academic papers, textbooks, and e-books in multiple languages.

Use Cases

• Benchmarking OCR model performance based on human-annotated quality scores.
• Comparing OCR text extraction accuracy across different document types mentioned in the description, such as academic papers and textbooks.
• Training or fine-tuning OCR quality assessment models using the provided human-annotated scores.
• Analyzing OCR error patterns in multilingual document processing.

Strengths

• Contains 1,000 document images, providing a substantial sample for evaluation.
• Includes human-annotated quality scores, offering a reliable ground truth for benchmarking.
• Features OCR text extracted by a specific, named model (Qwen2.5-VL-72B), enabling controlled comparisons.
• Covers diverse document types and multiple languages as described.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The full description is truncated, requiring a visit to the dataset page for complete details.

Provenance

Source

huggingface

Collection Method

OCR text extracted by Qwen2.5-VL-72B model, with human-annotated quality scores.

Freshness

Last updated 2025-10-17 08:10:31; freshness should be verified.