---
language:
- kmr
- ckb
- en
license: mit
library_name: transformers
tags:
- translation
- nllb-200
- badini
- kurmanji
- kurdish
- nlp
- seq2seq
pipeline_tag: translation
widget:
- text: How are you? Are you doing well?
  src_lang: eng_Latn
  tgt_lang: ckb_Arab
- text: I came from home.
  src_lang: eng_Latn
  tgt_lang: ckb_Arab
- text: ئەڤە قەلەمێ منە.
  src_lang: ckb_Arab
  tgt_lang: eng_Latn
---

# 🌟 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](https://instagram.com/arkan_it) |
| **Email** | [iarkanit@gmail.com](mailto:iarkanit@gmail.com) |
| **Official Website** | [tavbit.com](https://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

```python
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:

```bibtex
@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**.