TDoc-2.8M: 2.8 Million Multilingual Tampered Document Images

TDoc-2.8M is a large-scale dataset for tampered text detection and localization in document images, created by MohamedDhouib1. It contains approximately 2.8 million tampered document images, with English, French, and Chinese as the dominant languages. The full dataset size is about 7.44 TB and it was last updated on May 12, 2026.

Use Cases

• Train tampered text detection models based on the large-scale collection of manipulated document images.
• Develop multilingual document forgery localization systems based on images from multiple languages.
• Benchmark computer vision models for document integrity verification based on the dataset's scale and focus.

Strengths

• Approximately 2.8 million document images provide a substantial scale for model training.
• Multilingual coverage includes dominant languages English, French, and Chinese.
• The dataset is specifically designed for the focused task of tampered text detection and localization.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The specific distribution of images per language and the nature of tampering are not detailed in the provided metadata.
• The dataset's large size of 7.44 TB requires significant storage and computational resources for handling.

Provenance

Source

MohamedDhouib1 on Hugging Face.

Freshness

Last updated 2026-05-12 13:52:24; freshness should be verified.