MUDD - Misraj Unstructured Data Dump

4,758,338 rows of Arabic plain text, in 64 Parquet shards. Unlike MSDD it carries no images and no document structure: this is the translation arm of the pre-training data, Arabic rendered from English source material that had no Arabic counterpart worth crawling. The English side is SlimPajama-627B, the deduplicated open pre-training corpus.

Translation here is a coverage decision, not a volume exercise. The stage begins with a gap map (which domains are thinly represented in native Arabic web data) and selects high-value source material across technology, science, healthcare, law and culture. Each segment keeps its source identifier, original language, domain, licence and retrieval date. The translation is done by Mutarjim, a 1.5-billion-parameter decoder-only model built on Kuwain 1.5B, continued-pretrained on roughly ten billion tokens of Arabic-English pairs and then fine-tuned on about six million high-quality parallel pairs, weighted two to one in favour of Arabic-origin samples so the Arabic generator learns Arabic rhythm rather than imitating English structure. Output is accepted only after it clears language verification, a length-proportion check that catches truncation and padding, deduplication, and terminology validation.

Pre-training and continued pre-training where the goal is to add knowledge Arabic does not already carry: the domains a crawl of the Arabic web will not fill. It is machine-translated text and should be labelled as such in any mixture: the quality gates are there to hold down translation noise, not to make the output indistinguishable from originally-Arabic writing, and a model trained on too high a share of it will acquire a translated style.

Listed publicly on Hugging Face with access granted on request. The dataset card states no SPDX licence of its own, so the terms of the underlying SlimPajama-627B material and of the request itself are what govern use; the licence field on this page is deliberately empty rather than guessed.

The Mutarjim paper documents the translation model, its two-phase training and the Tarjama-25 benchmark released with it; the Kuwain paper documents the base model Mutarjim was built on. MSDD is the sibling dump from the web pipeline.