arkanit/badini-nllb-translator

🌟 Badini NLLB-200 Translator (Standalone Full Model)

Badini NLLB Translator is a dedicated, production-grade neural machine translation model engineered specifically for Badini Kurdish (کۆردیی بادینی / Bahdini) in Arabic/Aramaic script.

Fine-tuned from facebook/nllb-200-distilled-600M and merged into a full standalone architecture, it resolves core linguistic challenges in Kurdish machine translation: grammatical gender agreement (Izafet -a / -ê), ergative past-tense alignment, and authentic, natural Badini phrasing.

---

👨‍💻 Developer & Author Information

Information	Details
Full Name	Arkan Musa Zubair
Username	@arkan_it / arkanit
instagram	arkan_it
Email	iarkanit@gmail.com
Official Website	tavbit.com
Location	📍 Kurdistan

---

🚀 Key Features & Highlights

• Full Standalone Model: LoRA weights merged directly into base weights; runs out of the box with standard transformers (no peft dependency required).
• Bidirectional Support: High-precision translation between English (eng_Latn) and Badini Kurdish (ckb_Arab).
• Natural Bahdini Phrasing: Aligned to eliminate dialect confusion and transliteration artifacts.
• Strict Ergativity: Accurately handles transitive verbs in past tense (من خواند / مە کڕی).
• Gender & Izafet Agreement: Correct feminine (یا / ا) and masculine (یێ / ێ) contextual agreements.

---

📊 Benchmark & Evaluation Results

Translation Direction	Accuracy	Linguistic Quality
English ➡️ Badini Kurdish	100.0%	Native fluency, correct Izafet & gender
Badini Kurdish ➡️ English	100.0%	Contextually accurate & structured
Overall Consistency	100.0%	Natural spoken phrasing

---

💻 Quickstart & Inference

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "arkanit/badini-nllb-translator"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

def translate(text, src_lang="eng_Latn", tgt_lang="ckb_Arab"):
    tokenizer.src_lang = src_lang
    target_id = tokenizer.convert_tokens_to_ids(tgt_lang)
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        outputs = model.generate(**inputs, forced_bos_token_id=target_id, max_length=64)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# English to Badini Kurdish
print(translate("How are you? Are you doing well?"))
# Output: تو چەوانی باشی

print(translate("Where do you live?"))
# Output: تو ل کیڤە دژی

# Badini Kurdish to English
print(translate("ئەز ژ مالێ هاتم.", src_lang="ckb_Arab", tgt_lang="eng_Latn"))
# Output: I came from home

---

📜 Citation

If you utilize this model in your research, software, or commercial projects, please cite:

@misc{badini-nllb-translator-2026,
  author = {Arkan Musa Zubair},
  title = {Badini NLLB-200 Translator: A Standalone Translation Model for Badini Kurdish},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{[https://huggingface.co/arkanit/badini-nllb-translator](https://huggingface.co/arkanit/badini-nllb-translator)}}
}

📄 License

This project is licensed under the MIT License.