MoxhiMT 60 zh-vi

Chinese → Vietnamese Marian-style machine translation model, tuned for xianxia / web-novel text.

Intended Use

• Chinese → Vietnamese web novel / fiction translation
• Local or server inference
• Experimental release; review output for high-stakes / publication use

Model Details

• Architecture: Marian seq2seq (8 encoder + 2 decoder layers)
• Parameters: ~57M (d_model 576, ffn 2304)
• Tokenizer: SentencePiece source/target, joint ZH+VI, vocab 24k
• Suggested decoding: num_beams=4, max_length=512

Quick Start

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "DanVP/MoxhiMT-60"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

text = "他抬头看向远处的山门。"
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))

CTranslate2 (INT8)

A CTranslate2 INT8 build is included under ct2-int8/ for faster CPU inference.

import ctranslate2
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("DanVP/MoxhiMT-60")
translator = ctranslate2.Translator("ct2-int8", device="cpu", compute_type="int8")

text = "他抬头看向远处的山门。"
src = tok.convert_ids_to_tokens(tok(text, truncation=True, max_length=512).input_ids)
results = translator.translate_batch([src], beam_size=4, max_decoding_length=512)
print(tok.decode(tok.convert_tokens_to_ids(results[0].hypotheses[0]), skip_special_tokens=True))

Notes

• Prioritizes translation quality on xianxia / cultivation terminology.
• Trained from scratch with a custom SentencePiece-BPE 24k joint ZH+VI tokenizer.