Manhwa-NLLB-1.3B-EnAr-v2

Manhwa-NLLB-EnAr-v2

English to Arabic Manhwa/Webtoon Translation Model

What's New in V2?

This is a major upgrade from V1 with significant improvements in translation quality, model capacity, and domain expertise.

Exemple

---

Key Improvements

---

Translation Quality Comparison

Cultivation Terminology

Idioms & Expressions

Formal Greetings

---

🎯 Specialized Domain Coverage

This model excels at translating:

• Cultivation Systems: Qi, Dantian, Meridians, Realms, Breakthroughs
• Manhwa Idioms: "Courting death", "Frog in a well", "Mount Tai"
• Combat Terminology: Killing intent, Sword Intent, Techniques
• Character Titles: Young Master, Sect Leader, Supreme Elder
• System Prompts: Alerts, proficiency levels, skill notifications
• Arrogant Speech: Formal insults and threats common in cultivation stories

---

🚀 Quick Start

Installation

pip install torch transformers peft bitsandbytes accelerate

Basic Usage

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, BitsAndBytesConfig
from peft import PeftModel

# Configure 4-bit quantization for efficiency
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
)

# Load base model
base_model_id = "facebook/nllb-200-distilled-1.3B"
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForSeq2SeqLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "Youme21/Manhwa-NLLB-EnAr-v2")
model.eval()

# Translate
def translate(text):
    tokenizer.src_lang = "eng_Latn"
    inputs = tokenizer(text, return_tensors="pt").to(model.device)
    
    generated_tokens = model.generate(
        **inputs,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids("arb_Arab"),
        max_length=128,
        num_beams=5,
        early_stopping=True
    )
    
    return tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]

# Example
result = translate("You are courting death!")
print(result)  # أنت تتودد للموت!

---

📈 Training Details

Dataset

• Size: 122,683 curated translation pairs
• Split: 98% train (120,230) / 2% validation (2,453)
• Quality: Ultra-clean, manually curated manhwa-specific translations
• Domains: Cultivation, martial arts, system notifications, character dialogue

Training Configuration

• Base Model: facebook/nllb-200-distilled-1.3B
• Method: QLoRA (4-bit quantization)
• LoRA Rank: 16
• LoRA Alpha: 32
• Trainable Parameters: 23,592,960 (1.69% of total)
• Batch Size: 48 (effective)
• Training Steps: 2,505
• Final BLEU: 43.10

Hardware

• Trained on Google Colab (T4 GPU)
• Memory efficient through 4-bit quantization

---

Example Translations

examples = [
    "You have eyes but failed to recognize Mount Tai.",
    "This seat will personally teach you a lesson.",
    "I feel the Qi gathering in my Dantian.",
    "System Alert: Proficiency of 'Alchemy' has reached Master level."
]

Outputs:

• لديك عيون ولكن فشلت في التعرف على جبل تاي.
• هذا المقعد سيعلمك درسًا شخصيًا.
• أشعر بتجمع التشي في الدانتيان الخاص بي.
• تنبيه النظام: وصلت إجادة الكيمياء إلى مستوى الماجستير.

---

Performance Metrics

---

Version History

V2 (Current)

• Upgraded to NLLB-1.3B base model
• 122K training pairs with ultra-clean curation
• LoRA fine-tuning for efficiency
• BLEU score: 43.10
• Improved domain terminology

V1

• NLLB-600M base model
• ~80K training pairs
• Standard fine-tuning
• BLEU score: ~42.9

---

Use Cases

• Webtoon Translation: Translate Korean/Chinese webtoons to Arabic
• Novel Conversion: Adapt cultivation novels for Arabic readers
• Game Localization: RPG and cultivation game dialogue
• Fan Translations: Community-driven manhwa projects
• Cultural Adaptation: Preserve genre-specific terminology

---

Limitations

• Optimized for manhwa/cultivation genre (may not perform well on general text)
• Best with short-to-medium sentences (typical of dialogue bubbles)
• Requires specific terminology context for optimal results
• Arabic output may need post-editing for formal publications

---

Citation

If you use this model, please cite:

@misc{manhwa-nllb-enar-v2,
  author = {Youme21},
  title = {Manhwa-NLLB-EnAr-v2: English to Arabic Manhwa Translation},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Youme21/Manhwa-NLLB-EnAr-v2}
}

---

Contact & Feedback

• Issues: Open an issue on the model repository
• Improvements: Contributions and suggestions welcome!
• Community: Join discussions in the Community tab

---

📜 License

This model is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC-BY-NC-4.0).

You are free to:

• Share — copy and redistribute the material
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit and link to the license
• 🚫 NonCommercial — You may not use the material for commercial purposes

For commercial use, please contact the author.

Full license: https://creativecommons.org/licenses/by-nc/4.0/

---

🙏 Acknowledgments

• Meta AI for the NLLB model family
• HuggingFace for infrastructure and tools
• The manhwa translation community for inspiration and feedback

---

Built with ❤️ for the Arabic manhwa community