A lightweight BiLSTM that restores Arabic diacritics (tashkeel). Version 2 of TigreGotico/rawi.

Scored on the entire 817k-sentence held-out test split of TigreGotico/arabic_diacritized_text:

V2 is trained to abstain on unmarked positions (correct ignore_index masking over the padded loss), so it does not over-mark. INT8 quantization is lossless (2.30% DER, 2.5 MB).

On the corpus: it aggregates many public Arabic sources. rawi-v2 trained only on the train split and never on test.txt, though ~1.7% of test sentences also appear in train via near-duplicate aggregation — a smaller exposure than typical external baselines on the same set.

Embedding(236, 128) → 2-layer bidirectional LSTM(256) → Linear(512, 75). Normalization is NFD with symbols (Unicode So) dropped; the model predicts one of 75 NFD-based diacritic classes per base character (it also restores hamzas and the superscript alef).

text2tashkeel bundles this model and reproduces the exact NFD normalization and letter-only decode in pure Python (no PyTorch at inference).