quickmt-ja-en-v2

quickmt-ja-en-v2

A full fine-tune of quickmt/quickmt-ja-en, a fast and compact 200M-parameter Japanese → English translation model. Fine-tuned on approximately 31M sentence pairs from Yokii2/patchouli-jaen, a large-scale synthetic dataset of diverse Japanese web text (cc100-ja) translated by Mistral Small 4.

The model is exported in CTranslate2 format with 32k SentencePiece vocabularies, preserving the speed and format of the original.

Usage

Install the quickmt library first:

pip install git+https://github.com/quickmt/quickmt.git

Then load and run the model:

from quickmt import Translator
from huggingface_hub import snapshot_download

model_path = snapshot_download(
    "Yokii2/quickmt-ja-en-v2",
    ignore_patterns=["eole-model/*", "eole_model/*"]
)

mt = Translator(model_path, device="cpu")

sample_text = 'ノバスコシア州ハリファックスにあるダルハウジー大学医学部教授でカナダ糖尿病協会の臨床・科学部門の責任者を務めるエフード・ウル博士は、この研究はまだ初期段階にあるとして注意を促しました。'
translation = mt(sample_text, beam_size=5)

print("Original:   ", sample_text)
print("Translation:", translation)

Set device="cuda" for GPU inference. Use beam_size=1 for faster (but lower quality) output.

Translation Example

Benchmark Results

Evaluated on Yokii2/jmdict-ja-en-bench, a benchmark of 3,000 sentence pairs (1,000 per split) derived from JMDict vocabulary entries. Metrics: sacreBLEU, chrF2, and COMET-22 on 1,000 sentences per split. Time measured on an T4 (Google Colab) GPU at batch size 32.

Attribution

• Base model: quickmt/quickmt-ja-en by quickmt, licensed CC BY 4.0
• Training data: Yokii2/patchouli-jaen
• Benchmark: Yokii2/jmdict-ja-en-bench

This model is distributed under CC BY 4.0, in accordance with the base model license.