README
RealText-V2: A Large-Scale Multilingual Document Forgery Analysis Benchmark
💾 Dataset Description
RealText-V2 is a large-scale multilingual document benchmark dataset purpose-built for multilingual text image forgery analysis, pioneering in both scale and annotation depth.
Key Features
- 20K+ images: A large-scale benchmark, surpassing existing document forgery analysis datasets by orders of magnitude
- 6 languages: English, Chinese, Arabic, Thai, Malay, and Indonesian — spanning Latin, logographic, Arabic, and Thai script systems, each presenting unique forgery analysis challenges
- 6 domains: Finance, education, healthcare, live streaming, e-commerce, and natural scenes
- Multi-granularity forgery: Character-level, word-level, and semantic-level tampering
- Multi-source samples: Real-world and AIGC-synthesized forgery samples covering diverse generation pipelines
- Rich multi-task annotations: Pixel-level localization masks, tampering type labels, and expert-level natural language explanations
Competition Timeline
ACM MM 2026 MGC: GenText-Forensics: Challenge on Explainable Forensics and Adversarial Generation for Text-Centric Images
https://www.codabench.org/competitions/15805/
| Phase | Date |
|---|---|
| Competition Launch | April 17, 2026 |
| Training Data Release | April 20, 2026 |
| Evaluation Submission Opens | May 22, 2026 |
| Leaderboard Freeze | May 31, 2026 |
| Paper Submission Deadline | June 20, 2026 |
| ACM MM 2026, Rio de Janeiro | November 10–14, 2026 |
📊 Dataset Structure
RealText-V2/
├── train/
│ ├── image/ # Document images (.jpg for forged, .png for pristine)
│ │ ├── part000/ # Sharded at 1000 files per subdirectory
│ │ ├── part001/
│ │ └── ...
│ ├── mask/ # Binary tampering masks (forged only)
│ │ ├── part000/
│ │ └── ...
│ └── report/ # Structured forgery analysis reports (.md)
│ ├── part000/
│ ├── part001/
│ └── ...
├── doc_sample.png
└── metadata.parquet # Index file with sample metadata
Note: The test split is withheld for the ongoing ACM MM 2026 competition and will be released after the competition concludes.
Splits
| Split | Total | Black (Forged) | White (Pristine) |
|---|---|---|---|
| train | 13,500 | 7,500 | 6,000 |
Language Distribution (Train)
| Language | Code | Black (Forged) | White (Pristine) |
|---|---|---|---|
| English | en | 2,000 | 1,000 |
| Chinese | zh | 2,000 | 1,000 |
| Thai | th | 1,000 | 1,000 |
| Malay | ms | 1,000 | 1,000 |
| Indonesian | id | 1,000 | 1,000 |
| Arabic | ar | 500 | 1,000 |
📋 Data Fields
| Field | Description |
|---|---|
sample_id | Unique identifier (e.g., GenText_Forensic_00000000) |
language | Full language name |
language_code | ISO 639-1 code |
type | black (forged) or white (pristine) |
image_file | Image filename |
mask_file | Mask filename (empty for white samples) |
has_mask | Whether tampering mask exists |
report_file | Report filename |
report_text | Full report content |
📝 Report Format
Each report is a structured markdown document:
# FORGERY ANALYSIS REPORT
**[Conclusion]:** FORGED / PRISTINE
**[RISK_SCORE]:** 0-100
### ANOMALY_001: [type] ([location])
[GROUNDING]: [x1, y1, x2, y2]
[REASON]: [explanation text]
## SUMMARY
[summary text]
⚖️ License
This dataset is released under CC-BY-NC-4.0 for research purposes only.
🙏 Acknowledgments
RealText-V2 is created for the ACM MM 2026 competition on document forgery analysis.
Contributors
48 commits
Similar repos
README
RealText-V2: A Large-Scale Multilingual Document Forgery Analysis Benchmark
💾 Dataset Description
RealText-V2 is a large-scale multilingual document benchmark dataset purpose-built for multilingual text image forgery analysis, pioneering in both scale and annotation depth.
Key Features
- 20K+ images: A large-scale benchmark, surpassing existing document forgery analysis datasets by orders of magnitude
- 6 languages: English, Chinese, Arabic, Thai, Malay, and Indonesian — spanning Latin, logographic, Arabic, and Thai script systems, each presenting unique forgery analysis challenges
- 6 domains: Finance, education, healthcare, live streaming, e-commerce, and natural scenes
- Multi-granularity forgery: Character-level, word-level, and semantic-level tampering
- Multi-source samples: Real-world and AIGC-synthesized forgery samples covering diverse generation pipelines
- Rich multi-task annotations: Pixel-level localization masks, tampering type labels, and expert-level natural language explanations
Competition Timeline
ACM MM 2026 MGC: GenText-Forensics: Challenge on Explainable Forensics and Adversarial Generation for Text-Centric Images
https://www.codabench.org/competitions/15805/
| Phase | Date |
|---|---|
| Competition Launch | April 17, 2026 |
| Training Data Release | April 20, 2026 |
| Evaluation Submission Opens | May 22, 2026 |
| Leaderboard Freeze | May 31, 2026 |
| Paper Submission Deadline | June 20, 2026 |
| ACM MM 2026, Rio de Janeiro | November 10–14, 2026 |
📊 Dataset Structure
RealText-V2/
├── train/
│ ├── image/ # Document images (.jpg for forged, .png for pristine)
│ │ ├── part000/ # Sharded at 1000 files per subdirectory
│ │ ├── part001/
│ │ └── ...
│ ├── mask/ # Binary tampering masks (forged only)
│ │ ├── part000/
│ │ └── ...
│ └── report/ # Structured forgery analysis reports (.md)
│ ├── part000/
│ ├── part001/
│ └── ...
├── doc_sample.png
└── metadata.parquet # Index file with sample metadata
Note: The test split is withheld for the ongoing ACM MM 2026 competition and will be released after the competition concludes.
Splits
| Split | Total | Black (Forged) | White (Pristine) |
|---|---|---|---|
| train | 13,500 | 7,500 | 6,000 |
Language Distribution (Train)
| Language | Code | Black (Forged) | White (Pristine) |
|---|---|---|---|
| English | en | 2,000 | 1,000 |
| Chinese | zh | 2,000 | 1,000 |
| Thai | th | 1,000 | 1,000 |
| Malay | ms | 1,000 | 1,000 |
| Indonesian | id | 1,000 | 1,000 |
| Arabic | ar | 500 | 1,000 |
📋 Data Fields
| Field | Description |
|---|---|
sample_id | Unique identifier (e.g., GenText_Forensic_00000000) |
language | Full language name |
language_code | ISO 639-1 code |
type | black (forged) or white (pristine) |
image_file | Image filename |
mask_file | Mask filename (empty for white samples) |
has_mask | Whether tampering mask exists |
report_file | Report filename |
report_text | Full report content |
📝 Report Format
Each report is a structured markdown document:
# FORGERY ANALYSIS REPORT
**[Conclusion]:** FORGED / PRISTINE
**[RISK_SCORE]:** 0-100
### ANOMALY_001: [type] ([location])
[GROUNDING]: [x1, y1, x2, y2]
[REASON]: [explanation text]
## SUMMARY
[summary text]
⚖️ License
This dataset is released under CC-BY-NC-4.0 for research purposes only.
🙏 Acknowledgments
RealText-V2 is created for the ACM MM 2026 competition on document forgery analysis.
Similar repos
Linked in READMEs
Contributors
48 commits