DocSynth300K: 300,000 Document Layouts for Large-Scale Pre-training

DocSynth300K provides 300,000 document layout records for large-scale model pre-training, totaling 113GB in size. Released by researcher juliozhao in October 2024, the dataset is designed to enhance model performance in document layout analysis tasks. It is distributed in Parquet format and is compatible with high-performance data libraries like Polars and Dask.

Use Cases

• Pre-training transformer models using document layout labels to improve structural understanding
• Benchmarking layout engines against the 300,000 provided document structures
• Training OCR-agnostic models using the text content and spatial coordinates

Strengths

• 300,000 document records
• 113GB total data volume for large-scale pre-training
• Parquet format for efficient processing with Polars and Dask

Limitations

• High storage requirement of 113GB may be prohibitive for some users
• Synthetic nature of the data may not fully reflect the noise found in physical document scans
• Lack of explicit column documentation in the repository metadata

Provenance

Source

Arxiv paper 2410.12628

Collection Method

Synthetic generation

Freshness

Last updated in October 2024, corresponding with the release of Arxiv paper 2410.12628.