HY-MT1.5-1.8B-Trad-Chinese-ORPO

HY-MT1.5-1.8B-Trad-Chinese-ORPO

This model is a fine-tuned version of tencent/HY-MT1.5-1.8B specialized for English to Traditional Chinese translation.

It was trained using ORPO (Odds Ratio Preference Optimization) via the Unsloth library to enhance its ability to generate high-quality Traditional Chinese while explicitly rejecting Simplified Chinese characters and phrasing.

Model Details

• Base Model: tencent/HY-MT1.5-1.8B
• Training Method: ORPO (Odds Ratio Preference Optimization)
• Quantization: 4-bit (bitsandbytes)
• Language Pair: English -> Traditional Chinese
• Fine-tuning Tool: Unsloth & TRL

Training Data

The model was trained on 10,000 samples derived from the HuggingFaceFW/finetranslations (Mandarin subset).

• Chosen: Traditional Chinese translations (converted and verified using OpenCC).
• Rejected: Original translations containing Simplified Chinese or mixed scripts.
• Prompt: Translate the following segment into Traditional Chinese, without additional explanation.\n\n{English Text}

Performance Improvements

Through ORPO training, the model's preference for Traditional Chinese was significantly strengthened:

• Rewards Margin: Increased from ~0.07 to ~0.35 during training.
• Accuracy: Maintained near 100% in distinguishing preferred Traditional Chinese responses.

Usage

You can load this model using the unsloth library for fast inference:

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "yongjer/HY-MT1.5-1.8B-Trad-Chinese-ORPO",
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "user", "content": "Translate the following segment into Traditional Chinese, without additional explanation.\n\nArtificial Intelligence is transforming the world at an unprecedented pace."},
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

outputs = model.generate(input_ids=inputs, max_new_tokens=128)
print(tokenizer.batch_decode(outputs))

Acknowledgments

• Thanks to the Unsloth AI team for their efficient fine-tuning library.
• Original model by Tencent Hunyuan.