DanVP/MoxhiMT-30

Moxhi MT 30 zh-vi

Fast Chinese to Vietnamese Marian-style machine translation model, trained for web-novel / xianxia content.

Intended Use

• Chinese -> Vietnamese web novel / fiction translation (xianxia, modern, cross-domain).
• Fast local or server inference where a small model is preferred.
• Strong general / cross-domain coverage (sci-fi, historical, modern, mystery) while keeping xianxia / wuxia / classical register sharp.
• Experimental release; review output for high-stakes or publication use.

Model Details

• Architecture: Marian seq2seq (asymmetric 8 encoder + 2 decoder)
• Parameters: ~37M
• Tokenizer: SentencePiece joint source/target, 24k
• Suggested decoding: num_beams=4, max_length=512

Versions

Tag	Notes
v4.0 (current main)	New version.
v3.0	Pin with revision="v3.0".
v2.2	Pin with revision="v2.2".

Pin a specific version:

from transformers import AutoModelForSeq2SeqLM
model = AutoModelForSeq2SeqLM.from_pretrained("DanVP/MoxhiMT-30", revision="v3.0")

Quick Start

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "DanVP/MoxhiMT-30"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

Text = " hắn ngẩng đầu nhìn về phía xa xa sơn môn."
inputs = tok(text, return_tensors="pt", truncation=True, max_length=512)
out = model.generate(**inputs, max_length=512, num_beams=4)
print(tok.decode(out[0], skip_special_tokens=True))

Fast CPU Runtime

A CTranslate2 INT8 export is in ct2-int8_float32/ for ~3-5x faster CPU inference.

import ctranslate2
from pathlib import Path
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer

model_id = "DanVP/MoxhiMT-30"
model_path = Path(snapshot_download(model_id, allow_patterns=[
    "config.json", "source.spm", "target.spm", "vocab.json",
    "tokenizer_config.json", "ct2-int8_float32/*",
]))
tokenizer = AutoTokenizer.from_pretrained(model_path)
translator = ctranslate2.Translator(
    str(model_path / "ct2-int8_float32"),
    device="cpu", compute_type="int8_float32",
)

Training Data

Trained from scratch on a curated Chinese-Vietnamese parallel corpus covering xianxia, modern fiction, historical, sci-fi, and cross-domain web-novel content, with a research-grounded layer for idioms and classical-Chinese grammar, then a light preference-tuning (DPO) pass for xianxia/idiom sharpness.

Notes

• Prioritizes speed and small footprint.
• Known hard cases include rare proper nouns and highly domain-specific OOD terminology.
• For production usage, pair with reviewed glossary/guard layers where appropriate.

License

CC-BY-NC-4.0 (research / non-commercial use).