Loading datasets...
Release Date: 11/4/2026
Format: JPEG, JSONL, PARQUET, PDF, PNG, TIFF, WEBP
Size: 14.32 GB
The PashtoAI Vision Corpus, prepared by PashtoAI and the Digital Pashto Association, is a Pashto document-vision dataset for optical character recognition (OCR), handwritten text recognition and document understanding. The corpus has two parts. The labeled part contains 46,957 unique OCR image–text records, each pairing an image of Pashto text with its transcription. The raw part contains 30,182 unique files, including handwritten notes, notebook scans, educational documents, page images and PDFs, together about 16.6 GB before compression. It also includes 2,185 page images rendered as JPEG at 200 DPI. The collection combines real-world document images with synthetic OCR samples, covering a wide range of handwriting styles, fonts, layouts and image qualities. Labeled records include the original and Unicode-NFC-normalized Pashto text, image dimensions, technical quality indicators, duplicate flags, content-review indicators and deterministic train/validation/test splits. Every retained file has a SHA-256 checksum and structured provenance metadata. The dataset is intended for training and evaluating Pashto OCR and handwriting recognition models, document layout analysis, page segmentation, automated transcription and multimodal Pashto AI research. The handwritten and document materials were contributed by community members through the PashtoAI data platform (data.pashtoai.org), who consented to their open release. The dataset is released under the CC0 1.0 Universal Public Domain Dedication, allowing anyone to use, copy, modify, redistribute and build on it for research, educational, commercial and other purposes, to the extent of the rights held by the contributors and dataset creators. Limitations: the raw files are not all transcribed and are mainly suited to annotation, pre-training and evaluation work. Synthetic samples differ from real documents and should be weighted accordingly. Handwritten and scanned materials may contain names or other personal details; users must not attempt to identify individuals from them.
Released under the CC0 1.0 Universal Public Domain Dedication. The dataset owner confirms that the handwritten and document materials were contributed through the PashtoAI data platform (data.pashtoai.org) by community contributors who consented to their open release, including public redistribution, AI/ML training, research and commercial use. The dedication covers the rights held by the contributors and dataset creators. No additional dataset-specific restrictions are imposed.
You agree not to attempt to identify, contact or profile any contributor, or any person whose name or personal details may appear in the handwritten or scanned materials. You agree not to use this dataset to infer personal information about individuals.
Ethical Review
Materials were contributed voluntarily by community members through the PashtoAI data platform (data.pashtoai.org). The dataset owner confirms that contributors were informed of the purpose of the collection and consented to the open release of their submissions, including use for AI/ML. No independent ethics-board certification is asserted. Automated checks do not replace a full manual privacy review; content-review indicators are included in the metadata.
Intended Use
Pashto optical character recognition (OCR), handwritten text recognition, document layout analysis, page segmentation, automated transcription, OCR model training and evaluation, and multimodal Pashto document AI research.
Collection: The materials were contributed by community members through the PashtoAI data platform (data.pashtoai.org) and structured document-collection initiatives run by PashtoAI and the Digital Pashto Association. They include handwritten notebooks, scanned documents, educational materials, page images, PDFs and labeled OCR image–text pairs, alongside synthetic OCR samples.
Processing: Files went through integrity verification with SHA-256 checksums, exact-duplicate detection, image validation and dimension checks, and were organized into labeled and raw partitions. Text keeps the original Pashto alongside Unicode-NFC-normalized and whitespace-standardized versions. PDF pages were rendered as 2,185 JPEG images at 200 DPI.
Composition: 46,957 unique labeled OCR image–text records; 30,182 unique retained raw files (about 16.6 GB before compression); file formats HEIC, JPEG, PNG, TIFF, WEBP, PDF, JSONL and Parquet.
Limitations: Automated checks confirm file integrity and readability, not the accuracy of every transcription. Not all raw files are transcribed. Synthetic samples may not reflect real-world document conditions. Content-review indicators are included in the metadata, but a full manual privacy review of every file has not been completed.
