Loading datasets...
Release Date: 8/5/2026
Format: Parquet
Size: 107.99 MB
Share
This dataset contains 600,000 paired Urdu text samples for error correction tasks, where each entry pairs an "incorrect" text containing simulated character-level and typographic errors (repeated letters, garbled characters, OCR/typing-style noise) with its corresponding "correct" ground truth text. The ground truth Urdu sentences were created by Proxima AI, while the corresponding incorrect sentences were produced by applying controlled synthetic character-level, spelling, typographic, and OCR style transformations. The dataset is intended for training and evaluating Urdu text-correction and normalization models.
Licensing
Creative Commons Attribution Non Commercial Share Alike 4.0 International (CC-BY-NC-SA-4.0)
Restrictions/Special Constraints
This dataset is released under the CC BY-NC-SA 4.0 license for non-commercial research and educational use. The ground truth text was created in house by Proxima AI. Because Proxima AI owns the copyright to the original text, no third party redistribution permission was required. Users are also responsible for ensuring that their use of the dataset complies with applicable copyright law and the terms of the CC BY-NC-SA 4.0 license.
Forbidden Usage
The following uses are prohibited: commercial use except as permitted under the CC BY-NC-SA 4.0 license; redistribution without appropriate attribution; removal of required attribution or copyright notices; claiming ownership of the dataset; using the dataset to identify individuals or infer sensitive personal information; and distributing modified versions under terms incompatible with the CC BY-NC-SA 4.0 license.
Ethical Review
The ground-truth `correct` text was created in-house by Proxima AI. No third-party redistribution permission was required because Proxima AI owns the original content. The paired `incorrect` text was generated synthetically by applying controlled character-level, spelling, typographic, and OCR-style transformations. No personal, sensitive, or user-generated data is included. The errors were introduced through an automated script rather than collected from real users, so no human-subjects review was required.
Intended Use
This dataset is intended for training and evaluating Urdu grammatical and spelling error correction models, and for research on text normalization and robustness to OCR/typing style noise in low-resource NLP settings.
Task: Urdu Grammatical Error Correction
Language: Urdu (ur)
Total records: 600,000
Format: PARQUET
Encoding: UTF-8
| Field | Description |
|---|---|
| input_text | Urdu sentence containing synthetic spelling or typographical errors |
| correct_text | Correct ground-truth Urdu sentence |
| input_text | correct_text |
|---|---|
| پاکستان ایک خووبصورت ملک ہے۔ | پاکستان ایک خوبصورت ملک ہے۔ |
| مجھے اردو زبان بہت پسندد ہے۔ | مجھے اردو زبان بہت پسند ہے۔ |
| وہ اسکول جاا رہا ہے۔ | وہ اسکول جا رہا ہے۔ |
Create the ground-truth Urdu sentences in-house at Proxima AI.
Normalize Unicode and formatting.
Remove duplicate records.
Generate synthetic spelling and OCR-style errors.
Validate sentence alignment.
Export the final dataset in PARQUET format.
Errors are synthetically generated and may not cover every type of real-world writing error.
OCR-style noise is simulated rather than produced by a specific OCR engine.
Performance may vary on informal text, handwriting, or historical documents.
