AraToken-Qwen3-1.7B-LEP
Qwen3-1.7B-Base adapted to Arabic with LEP. The 130,890 new embedding rows were initialized with the hybrid estimate ½(compositional + analogical). The model was trained for 2,000 steps (196M tokens, batch 96 × 1,024), with the original embedding rows frozen and transformer layers 24–27 trainable.
Paper: AraToken: Optimizing Arabic Tokenization with Normalization Pipeline and Language Extension for Qwen3. Code: https://github.com/mariklolik/Aratoken.
| Model | Arabic BPC ↓ |
|---|---|
| Qwen3-1.7B-Base | 1.3313 |
| CPT with the original tokenizer (mariklolik/AraToken-Qwen3-1.7B-CPT) | 1.2672 |
| this model | 1.2598 |
The difference to CPT is significant at p < 0.001 (paired bootstrap over test documents).
BPC is bits per character on the first 1,500 documents of the held-out test split of
mariklolik/AraToken-FineWeb2-HQ-ar.
It compares models with different vocabularies on the same characters.
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "mariklolik/AraToken-Qwen3-1.7B-LEP"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="bfloat16")
The tokenizer carries the AraToken normalizer: NFKC, Tatweel removal, Western digits, Latin punctuation and diacritics removal. Raw Arabic text can therefore be passed to it directly.
Citation
@article{kashirskiy2025aratoken,
title = {AraToken: Optimizing Arabic Tokenization with Normalization Pipeline and Language Extension for Qwen3},
author = {Kashirskiy, Mark and Lipinski, Artiom and Makarov, Ilya},
journal = {arXiv preprint arXiv:2512.18399},
year = {2025}
}