08

Machine Translation

8.1 Neural Machine Translation

jparacrawl-finetune (MorinoseiMorizo/jparacrawl-finetune): Usage examples for the JParaCrawl pretrained NMT model. Stars: 104. JParaCrawl is a large-scale Japanese-English parallel corpus and an important resource for machine translation research.

JASS (Mao-KU/JASS): Proposes Japanese-specific Sequence-to-Sequence pretraining. Presented at LREC 2020. Stars: 16.

plamo-translate-cli (pfnet/plamo-translate-cli): A translation CLI using Preferred Networks' plamo-2-translate model. Stars: 351, with high popularity. Can be run locally.

jparacrawl-finetune
jparacrawl-finetune (MorinoseiMorizo/jparacrawl-finetune): Usage examples for the JParaCrawl pretrained NMT model. Stars: 104. JParaCrawl is a large-scale Japanese-English parallel corpus and an important resource for machine translation research.
104
JASS
JASS (Mao-KU/JASS): Proposes Japanese-specific Sequence-to-Sequence pretraining. Presented at LREC 2020. Stars: 16.
16
plamo-translate-cli
plamo-translate-cli (pfnet/plamo-translate-cli): A translation CLI using Preferred Networks' plamo-2-translate model. Stars: 351, with high popularity. Can be run locally.
351

8.2 Translation Tools

FuguMT-related tools:

  • tra-fugu: A Japanese-English and English-Japanese translation tool using FuguMT
  • fugumt: A FuguMT translation environment; also supports PDF translation

pdf-translator (discus0434/pdf-translator): A tool that translates English PDFs into Japanese while preserving the original layout. Stars: 347.

pdf-translator
pdf-translator (discus0434/pdf-translator): A tool that translates English PDFs into Japanese while preserving the original layout. Stars: 347.
347

8.3 Translation Evaluation

PheMT (cl-tohoku/PheMT): A phenomenon-based evaluation dataset for the robustness of Japanese-English machine translation. Contains additional annotations for four linguistic phenomena: proper nouns, abbreviations, colloquial expressions, and variants (COLING 2020). Stars: 20.

PheMT
PheMT (cl-tohoku/PheMT): A phenomenon-based evaluation dataset for the robustness of Japanese-English machine translation. Contains additional annotations for four linguistic phenomena: proper nouns, abbreviations, colloquial expressions, and variants (COLING 2020). Stars: 20.
20

Translation Models on Hugging Face

Model HF DL Likes Description
opus-mt-ja-en (Helsinki-NLP) 37k 77 Trained on Opus corpus; BLEU 41.7
vntl-llama3-8b-v2-gguf (lmg-anon) 20k 17 Visual novel Japanese to English translation
Sugoi-14B-Ultra-GGUF (sugoitoolkit) 10k 20 BLEU 21.38; strong on RPG-Maker text
LFM2-350M-ENJP-MT-GGUF (LiquidAI) 5k 45 Lightweight 350M-parameter translation
plamo-2-translate (pfnet) 2k 130 PFN large-scale translation model

vntl-llama3-8b-v2-gguf fell sharply from 714k to 20k downloads (#40 across all Japanese NLP models on Hugging Face), and the top of the translation category has passed to Helsinki-NLP's opus-mt-ja-en (37k, 77 Likes). vntl-llama3 is a LLaMA 3-based model optimized for translating Japanese visual novels into English, reflecting the demand for cultural content translation.

Sugoi-14B-Ultra-GGUF (sugoitoolkit) records 10k weekly downloads with BLEU 21.38, placing 60th overall on Hugging Face. With strong performance on RPG-Maker-style bracketed text and JSON output, it is widely adopted for game translation and content localization.

plamo-2-translate (Preferred Networks): Highly regarded with 130 Likes, and is a large language model designed specifically for translation tasks.


opus-mt-ja-en (Helsinki-NLP)
Trained on Opus corpus; BLEU 41.7
37k 77
vntl-llama3-8b-v2-gguf (lmg-anon)
Visual novel Japanese to English translation
20k 17
Sugoi-14B-Ultra-GGUF (sugoitoolkit)
BLEU 21.38; strong on RPG-Maker text
10k 20
LFM2-350M-ENJP-MT-GGUF (LiquidAI)
Lightweight 350M-parameter translation
5k 45
plamo-2-translate (pfnet)
PFN large-scale translation model
2k 130
plamo-2-translate
plamo-2-translate (Preferred Networks): Highly regarded with 130 Likes, and is a large language model designed specifically for translation tasks.