Sadeed Tashkeela

1,042,698 training rows and 2,485 test rows of diacritized Arabic, 1.43 GB in Parquet. Each row carries three fields: the source filename, the undiacritized input, and the fully diacritized output. The test split is held separate in the published files, so a fine-tune and its own held-out evaluation come out of the same download.

The corpus is derived from Tashkeela, the standard large Arabic diacritized collection, put through the cleaning and normalisation pipeline built for the Sadeed paper. That work exists because diacritization data in the wild is inconsistent: the same word appears with partial, full and absent diacritics across sources, and a model fine-tuned on the mixture learns the inconsistency. The pipeline normalises the diacritic forms, removes the entries that cannot be reconciled, and keeps the input and output strictly aligned so the task stays a character-faithful restoration rather than a rewrite.

Fine-tuning and evaluating Arabic diacritization models: the use it was built for. Sadeed itself is a fine-tune of Kuwain 1.5B on exactly this kind of curated diacritized data. Downstream, accurate diacritization matters most to Arabic text-to-speech, to machine translation where an unvowelled word is ambiguous, and to language-learning tools. For publishable evaluation numbers the SadeedDiac-25 benchmark is the fairer instrument: it was built precisely because existing diacritization benchmarks were too narrow in genre and complexity.

Listed publicly on Hugging Face with access granted on request. The card declares no SPDX licence of its own, so use is governed by the terms of the underlying Tashkeela material and of the access request; the licence field on this page is left empty rather than guessed.

The Sadeed paper documents both the cleaning pipeline behind this corpus and the model trained on it, and introduces SadeedDiac-25, the benchmark released to evaluate diacritization across genres and complexity levels. Kuwain 1.5B is the base model Sadeed was adapted from.