dictabert-large-char-menaked

DictaBERT-large-char-menaked: An open-source BERT-based model for adding diacritiziation marks ("nikud") to Hebrew texts

This model is a fine-tuned version of DictaBERT-large-char, dedicated to the task of adding nikud (diacritics) to Hebrew text.

The model was trained on a corpus of modern Hebrew texts manually diacritized by linguistic experts. As of 2025-03, this model provides SOTA performance on all modern Hebrew vocalization benchmarks as compared to all other open-source alternatives, as well as when compared with commercial generative LLMs.

Note: this model is trained to handle a wide variety of genres of modern Hebrew prose. However, it is not intended for earlier layers of Hebrew (e.g. Biblical, Rabbinic, Premodern), nor for poetic texts.

Sample usage:

from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('dicta-il/dictabert-large-char-menaked')
model = AutoModel.from_pretrained('dicta-il/dictabert-large-char-menaked', trust_remote_code=True)

model.eval()

sentence = 'בשנת 1948 השלים אפרים קישון את לימודיו בפיסול מתכת ובתולדות האמנות והחל לפרסם מאמרים הומוריסטיים'
print(model.predict([sentence], tokenizer))

Output:

['בִּשְׁנַת 1948 הִשְׁלִים אֶפְרַיִם קִישׁוֹן אֶת לִמּוּדָיו בְּפִסּוּל מַתֶּכֶת וּבְתוֹלְדוֹת הָאׇמָּנוּת וְהֵחֵל לְפַרְסֵם מַאֲמָרִים הוּמוֹרִיסְטִיִּים']

Matres Lectionis (אימות קריאה)

As can be seen, the predict method automatically removed all the matres-lectionis (אימות קריאה). If you wish to keep them in, you can specify that to the predict function:

print(model.predict([sentence], tokenizer, mark_matres_lectionis = '*'))

Output:

['בִּשְׁנַת 1948 הִשְׁלִים אֶפְרַיִם קִישׁוֹן אֶת לִי*מּוּדָיו בְּפִי*סּוּל מַתֶּכֶת וּבְתוֹלְדוֹת הָאׇמָּנוּת וְהֵחֵל לְפַרְסֵם מַאֲמָרִים הוּמוֹרִיסְטִיִּים']

Community Project

A third-party project, dicta-onnx, offers a lightweight ONNX-based tool built on top of our model for adding Hebrew diacritics. We're not affiliated, but it's a cool and practical application you might find useful.

License

Shield:

This work is licensed under a Creative Commons Attribution 4.0 International License.