mariklolik/AraToken-Qwen3-0.6B-CPT

AraToken-Qwen3-0.6B-CPT

The continued-pretraining (CPT) baseline of the AraToken paper. Qwen3-0.6B-Base keeps its original tokenizer and is trained on the same 500M-token FineWeb2-HQ Arabic budget as mariklolik/AraToken-Qwen3-0.6B-LEP: 5,086 steps, transformer layers 24–27 trainable.

Paper: AraToken: Optimizing Arabic Tokenization with Normalization Pipeline and Language Extension for Qwen3. Code: https://github.com/mariklolik/Aratoken.

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-Qwen3-0.6B-CPT"
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}
}