Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild
Paper • 2605.22064 • Published • 5
feishengmt
Chinese-to-English translation of wuxia and xianxia fiction. The model is a fine-tune of tencent/Hy-MT2-1.8B (LoRA, merged into the weights), trained on wuxia text. It is built mainly for English; a short second stage adapted it so that it also answers in French, Portuguese, Spanish, Indonesian, Japanese and Korean when one of them is asked for.
What it learned beyond the base model:
- The vocabulary and conventions of the genre: cultivation terms, honorifics (Senior / Junior Brother ...), recurring names.
- To work on paragraphs, not on short pieces. It writes the quotation marks by itself, also when the Chinese uses
「」or Traditional characters, when a quote runs over several paragraphs, and when the input starts or ends in the middle of a quote.
Version 1.0. It is a modified version of Hy-MT2-1.8B and is not endorsed by Tencent. A quantized GGUF version (1.1 GB) is at hsilvosa/feishengmt-GGUF and a WebGPU (ONNX) version for the browser is at hsilvosa/feishengmt-ONNX.
Use it with the included pipeline
The model works best through the pipeline in the pipeline/ folder of this repository. It cuts the text into windows of
whole paragraphs (up to 400 characters, or one paragraph when it is longer), adds the glossary entries found in each
window, checks that the answer has the same paragraphs and the same quote marks as the source, and translates again in
smaller parts when it does not.
import sys
from huggingface_hub import snapshot_download
path = snapshot_download("hsilvosa/feishengmt")
sys.path.insert(0, f"{path}/pipeline")
from webnovel_mt.translate import hf_backend
from webnovel_mt.windows import translate_document_windows
backend = hf_backend(path) # needs transformers>=5.6 and torch
chinese = open("chapter.txt", encoding="utf-8").read()
print(translate_document_windows(chinese, backend))
Without the pipeline, use this prompt. The language name goes in Chinese or in English (英语 or English).
将以下文本翻译为English，注意只需要输出翻译后的结果，不要额外解释：
<Chinese text: one or more paragraphs, up to about 400 characters>
Other languages
Replace the language name in the prompt: 法语 (French), 葡萄牙语 (Portuguese), 西班牙语 (Spanish), 印尼语
(Indonesian), 日语 (Japanese), 韩语 (Korean). The pipeline above is for English only: its checks of quote marks follow
English conventions and its glossary is English. For the other languages, send one paragraph or a short window at a time
and check the result yourself.
Results
All numbers are one run with greedy decoding on windows of chapters that were held out from training. Reference translations are from a human-made corpus; they follow their own conventions, so chrF is only a rough guide.
Chinese to English (pipeline, windows mode), two held-out sets from different novels, chrF and the share of answers whose quote marks are those of the reference. The variants perturb the Chinese source in the ways described above.
| Variant of the source | Set 1 (30 windows) chrF / quotes | Set 2 (40 windows, 28 for the split quote) chrF / quotes |
|---|---|---|
| Clean | 63.0 / 100 % | 47.9 / 97.5 % |
Corner quotes 「」 |
62.9 / 100 % | 48.1 / 95 % |
| Quote cut across paragraphs | 63.8 / 100 % | 48.3 / 96 % |
| Traditional characters | 63.0 / 100 % | 47.1 / 97.5 % |
Traditional and 「」 |
62.9 / 100 % | 47.1 / 95 % |
For comparison, the first version of this model (v0.1) with 250-character pieces scored chrF 44.8 on set 2, and
reproduced the quote marks of a quote split across paragraphs in none of the windows (against 96 % here). On a small check
with excerpts of older Chinese novels, 26 of 38 paragraphs with quotes came out with the same number of quote marks as
the original (19 for v0.1); most of the other 12 come from typos in the source and from book titles in 《》.
Other languages, 60 windows (30 from the novels of the training corpus, 30 from older novels), no references, so only two things were measured: whether the answer is in the requested language and whether it is as close in meaning to the source as the base model's answer (LaBSE similarity).
| Language | Answer in the requested language (name given in Chinese) | (name given in English, 20 windows) |
|---|---|---|
| French | 98 % | 100 % |
| Portuguese | 97 % | 95 % |
| Spanish | 97 % | 100 % |
| Indonesian | 97 % | 100 % |
| Japanese | 98 % | 85 % |
| Korean | 100 % | 100 % |
Before the second stage the figures were 48 to 92 % with the name in Chinese and 0 to 20 % with the name in English (Spanish was not measured): the model answered in English or mixed languages. The LaBSE similarity is the same as the base model's (differences of -0.02 to +0.01).
Training
- Base: Hy-MT2-1.8B. LoRA r=16, alpha=32, all attention and MLP projections, bf16, one RTX 3060.
- Stage 1, English. About 23,000 examples (7.4 million tokens) of Chinese paragraph windows with their English
translation, from a Chinese-English wuxia and xianxia corpus. The Chinese side was perturbed at random (Traditional
characters,
「」quotes, cut borders) while the English side kept clean double quotes, so that the model learns to write the quotes by itself. A glossary (honorifics, cultivation vocabulary, recurring names) is in the prompt during training. Learning rate 2e-4, one epoch, 715 steps, 2 h 12 min. - Stage 2, other languages. About 1,000 examples in the six other languages, made by the stage 1 model itself from Chinese training windows and kept only if they have the same paragraphs, are in the right language, have no English or Chinese left over and have a LaBSE similarity of at least 0.72 with the source. Training continued from stage 1 with these examples mixed with 1,500 English windows and 300 sentences, learning rate 1e-4, 99 steps (18 minutes). The English results above are after this stage.
- The corpus and the generated examples are not redistributed.
Limitations
- English is the priority. The other languages were only taught to follow the request, and their quality is about the one of the base model: they were made by this model and filtered for form, not for meaning. Names and some terms come out wrong or stay in English. Portuguese had the fewest training examples.
- Stage 2 cost a little in English: -0.4 chrF on average on set 1 and no change on set 2 (one run; the drop is within the noise of such a run, but it is negative in all five variants of set 1).
- Domain: wuxia and xianxia. Modern text, technical text and literary or classical Chinese were not trained or measured, apart from the small check above. The glossary encodes one convention (Senior/Junior); other translators use Martial Brother or Elder Brother.
- Small evaluation (two sets of 30 and 40 windows, one run). Differences below about 1 chrF point are noise. The non-English results have no reference translations.
- Known errors: dropped or repeated sentences in very long dialogues, characters left in Chinese when a name is rare, and a quote mark that is lost or added when the source has a typo in its quotes.
- A model of 1.8 billion parameters: do not use it where an error matters.
License and attribution
Apache-2.0. Based on Hy-MT2 by Tencent (Apache-2.0); see NOTICE. Files were modified (fine-tuning, weight merge).
@misc{zheng2026hymt2,
title={Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild},
author={Mao Zheng and Zheng Li and Tao Chen and Bo Lv and Mingrui Sun and Mingyang Song and Jinlong Song and Hong Huang and Decheng Wu and Hai Wang and Yifan Song and Yanfeng Chen and Guanwei Zhang},
year={2026},
eprint={2605.22064},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
- Downloads last month
- 18