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

Japanese	霧雨魔理沙はミニ八卦炉を構え、強力なマスタースパークを放った。
Fine-Tuned	Kirisame Marisa set up a mini eight-ring furnace and unleashed a powerful Master Spark.
Base model	Drizzle Ma Risa set up a mini-eight furnace and unleashed a powerful master spark.

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.

Model	Path	Benchmark	BLEU	chrF2	COMET-22	Time (s)
Fine-Tuned	Yokii2/quickmt-ja-en-v2	short	40.68	61.71	87.50	1.60
Fine-Tuned	Yokii2/quickmt-ja-en-v2	medium	39.95	65.36	89.86	2.96
Fine-Tuned	Yokii2/quickmt-ja-en-v2	long	37.52	64.64	88.38	4.05
CAT-Translate 0.8B	cyberagent/CAT-Translate-0.8b	short	35.77	57.52	87.02	306.76
CAT-Translate 0.8B	cyberagent/CAT-Translate-0.8b	medium	34.30	60.73	88.82	768.06
CAT-Translate 0.8B	cyberagent/CAT-Translate-0.8b	long	30.14	59.40	87.40	1518.90
Base	quickmt/quickmt-ja-en	short	33.45	53.83	84.35	1.57
Base	quickmt/quickmt-ja-en	medium	33.88	59.42	87.81	3.49
Base	quickmt/quickmt-ja-en	long	31.82	59.87	86.75	3.71
LiquidAI 350M	LiquidAI/LFM2-350M-ENJP-MT	short	27.82	53.65	84.39	32.06
LiquidAI 350M	LiquidAI/LFM2-350M-ENJP-MT	medium	33.27	60.44	88.62	76.74
LiquidAI 350M	LiquidAI/LFM2-350M-ENJP-MT	long	32.02	60.95	87.51	67.92
OPUS-MT	Helsinki-NLP/opus-mt-ja-en	short	25.91	46.92	80.42	11.52
OPUS-MT	Helsinki-NLP/opus-mt-ja-en	medium	22.15	49.13	82.84	18.01
OPUS-MT	Helsinki-NLP/opus-mt-ja-en	long	20.26	48.65	81.15	26.78

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.