opus-mt-en-vi-novel-finetuned

Fine-tuned English → Vietnamese translation model optimized for novels, literature, and narrative text.

Fine-tuned from Helsinki-NLP/opus-mt-en-vi (MarianMT) for 1 epoch on ~328K parallel EN-VI novel sentence pairs.

Results Summary

Full Validation Set (17,272 samples)

Model	BLEU	ChrF++
Fine-tuned	29.29	49.13
Baseline (Helsinki-NLP/opus-mt-en-vi)	3.73	20.87

Improvement over baseline

• BLEU: 8.2x (3.53 → 29.01 on sanity set)
• ChrF++: 2.4x (20.50 → 49.07 on sanity set)

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("richardadam/opus-mt-en-vi-novel-finetuned")
tokenizer = AutoTokenizer.from_pretrained("richardadam/opus-mt-en-vi-novel-finetuned")

text = "He looked at the door but did not dare to open it."
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, num_beams=4)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# Hắn nhìn cánh cửa, nhưng không dám mở ra.

Training

• Base: Helsinki-NLP/opus-mt-en-vi
• Epochs: 1
• Batch size: 16
• GPU: NVIDIA L4 (22.5 GB)
• Runtime: ~3h15m

Limitations

• Single-epoch fine-tune — may benefit from more training
• Novel domain only — may underperform on technical/formal text
• Baseline model performs very poorly on novels (BLEU 3.53), confirming domain-specific fine-tuning is essential

Citation

@misc{opus-mt-en-vi-novel-finetuned,
  author = {richardadam},
  title = {opus-mt-en-vi-novel-finetuned},
  year = 2026,
  publisher = {Hugging Face},
  howpublished = {huggingface.co/richardadam/opus-mt-en-vi-novel-finetuned},
}