Datasets

A 2-trillion-token proprietary Arabic corpus - the largest clean dataset built specifically for Arabic LLM pre-training. It underlies Kawn Lab's foundation models and is not publicly released.

35B+ curated Arabic tokens released publicly, growing regularly, so other researchers can build on the same data foundation rather than starting from scratch.

An open pipeline for constructing large-scale structured Arabic multimodal corpora from Common Crawl, preserving document structure in Markdown format rather than flattening it into noise.

A large-scale Arabic multimodal dataset created with the Wasm pipeline, extracted and filtered from Common Crawl dumps. It uniquely preserves the structural integrity of web content by emitting Markdown, addressing the shortage of high-quality structured multimodal data for Arabic.

A large-scale Arabic plain-text dataset of 4,758,338 rows, translated from SlimPajama-627B for pre-training Arabic language models. Each entry keeps its UUID from the original source.

A large, high-quality Arabic diacritized corpus for training and evaluating Arabic diacritization models. The training split is built from a cleaned version of the Tashkeela corpus; the test split is a refined version of the Fadel Tashkeela test set.

A large-scale Arabic multimodal dataset of 100 million Arabic image captions, generated with the Mutarjim translation model. It exists to close the gap in Arabic multimodal resources so Arabic vision-language systems can be trained at all.

A reviewed and corrected version of the KITAB-Bench PDF-to-Markdown subset for Arabic document OCR evaluation. Ground-truth errors - hallucinated text, missing page numbers, omitted small-font text - were fixed and formatting standardised, while the original task and schema are kept so it remains a drop-in evaluation set.