yennguyen45/vp2vi-polish-web

vp2vi (ONNX, INT8, web build)

Browser-ready build of DanVP/vp2vi, a MarianMT model that polishes VietPhrase/QT-style Chinese web-novel drafts into more natural Vietnamese (Vietnamese → Vietnamese, not a Chinese→Vietnamese translator).

All credit for training the base model goes to the original author. This repo only adds:

• ONNX export with KV-cache (encoder_model_quantized.onnx + decoder_model_merged_quantized.onnx)
• Dynamic INT8 quantization (~4x smaller than fp32)
• A hand-built tokenizer.json fast-tokenizer (the original model only ships slow sentencepiece files, which transformers.js cannot load directly) — verified to produce byte-identical token ids to the original slow MarianTokenizer.

Shares the same 24k joint ZH+VI vocabulary as yennguyen45/hachimimt-60-zh-vi-web, so the two are meant to be used as a pipeline: HachimiMT translates zh→vi, then this model polishes the result.

Usage (transformers.js)

import { pipeline } from '@huggingface/transformers';

const polish = await pipeline('translation', 'yennguyen45/vp2vi-polish-web', { dtype: 'q8' });
const out = await polish('Hắn chậm rãi ngẩng đầu, trong mắt lóe lên một tia hàn mang.');
console.log(out[0].translation_text);

License

CC-BY-4.0, inherited from the base model. Please credit the original author (DanVP) when using this model.