This is a custom-trained Marian-based translation model for src→dst. It was fine-tuned from Helsinki-NLP/opus-mt-zh-en and converted to MNN format via the mnn-opus-mt-toolkit.

Model Details

Parameter	Value
Architecture	MarianMT (encoder-decoder Transformer)
Source language	src
Target language	dst
Base model	`Helsinki-NLP/opus-mt-zh-en`

Usage

With Hugging Face Transformers

python

from transformers import AutoModelForSeq2SeqLM, MarianTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("Hosstia/opus-mt-zh-ru-32k")
tokenizer = MarianTokenizer.from_pretrained("Hosstia/opus-mt-zh-ru-32k")

text = "Your source text here"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs)
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)

With MNN (on-device Android)

Convert to MNN format using the mnn-opus-mt-toolkit:

bash

./run_all.sh --hf-user <username> --src src --dst dst --format fp16 \
    --model-repo Hosstia/opus-mt-zh-ru-32k --tokenizer-repo Hosstia/opus-mt-zh-ru-32k

License

Apache License 2.0 (model weights); see the toolkit repository for details.

Acknowledgments

Helsinki-NLP for the original OPUS-MT models
Marian NMT for the training framework
Alibaba MNN for the on-device inference engine
Hugging Face for model hosting and the Transformers library