Translate12M - Multilingual Translation Models

Collection of 12 LoRA fine-tuned translation models based on NLLB-200-distilled-600M.

🌐 Models

Direction	BLEU	Size (MB)	Path
EN→TH	99.53	143.5	eng_Latn_to_tha_Thai
TH→EN	99.67	143.5	tha_Thai_to_eng_Latn
EN→JA	98.66	143.5	eng_Latn_to_jpn_Jpan
JA→EN	99.55	143.5	jpn_Jpan_to_eng_Latn
EN→HI	99.70	143.5	eng_Latn_to_hin_Deva
HI→EN	99.84	143.5	hin_Deva_to_eng_Latn
EN→RU	99.52	143.5	eng_Latn_to_rus_Cyrl
RU→EN	99.38	143.5	rus_Cyrl_to_eng_Latn
EN→MS	99.37	143.5	eng_Latn_to_zsm_Latn
MS→EN	99.74	143.5	zsm_Latn_to_eng_Latn
EN→ZH	98.62	143.5	eng_Latn_to_zho_Hans
ZH→EN	99.13	143.5	zho_Hans_to_eng_Latn

🚀 Quick Start

from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# Example: EN→TH
direction = 'eng_Latn_to_tha_Thai'

config = PeftConfig.from_pretrained('wasawat/Translate12M', subfolder=direction)
base = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path)
model = PeftModel.from_pretrained(base, 'wasawat/Translate12M', subfolder=direction)
tokenizer = AutoTokenizer.from_pretrained('wasawat/Translate12M', subfolder=direction)

# Set languages
tokenizer.src_lang = 'eng_Latn'
tokenizer.tgt_lang = 'tha_Thai'

# Translate
text = 'Hello, how are you?'
inputs = tokenizer(text, return_tensors='pt')
outputs = model.generate(**inputs, max_length=128)
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translation)

📊 Details

• Base Model: NLLB-200-distilled-600M
• Method: LoRA fine-tuning
• LoRA Rank: 128
• LoRA Alpha: 256
• Training: ~4,250 samples per direction
• Epochs: 10
• Hardware: Tesla T4 GPU

📜 License

CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0)

• ✅ Free for research/education
• ❌ No commercial use
• ✅ Attribution required

👤 Author

wasawat

Repository: https://huggingface.co/wasawat/Translate12M