---
license: cc-by-4.0
language:
  - vi
base_model: DanVP/MoxhiMT-30-QT
library_name: transformers
pipeline_tag: translation
tags:
  - marian
  - vietnamese
  - text-rewriting
  - post-editing
  - web-novel
---

# vp2vi — VietPhrase/Convert to Natural Vietnamese

**vp2vi** is a Vietnamese-to-Vietnamese text-rewriting model for polishing
VietPhrase/QT-style Chinese web-novel drafts into more natural Vietnamese. It
is a 36.5M-trainable-parameter Marian model with 8 encoder layers, 2 decoder
layers, a 448-dimensional hidden size, and a shared 24,000-token vocabulary.

The model continues from
[DanVP/MoxhiMT-30-QT](https://huggingface.co/DanVP/MoxhiMT-30-QT), retaining
its full tokenizer, vocabulary, encoder, and two-layer decoder initialization.

## Quick start

```python
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "DanVP/vp2vi"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

draft = "Hắn đối với chuyện này cũng không có biện pháp nào."
inputs = tokenizer(draft, return_tensors="pt", add_special_tokens=False)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=False,
        num_beams=1,
        max_new_tokens=224,
    )

print(tokenizer.decode(output_ids[0], skip_special_tokens=True))
```

This is the standard Transformers compatibility path. Plain `generate()` can
still mishandle digits, repetition, or rare names; the
[hosted browser application](https://huggingface.co/spaces/DanVP/vp2vi) uses
additional decoding constraints. Its runtime is distributed separately and is
not part of this model repository.

## Intended use

- Sentence- or line-level polishing of Vietnamese VietPhrase/QT drafts from
  xianxia, xuanhuan, and related web-novel genres.
- Local reading tools, post-editing experiments, and further fine-tuning.

This is not a Chinese-to-Vietnamese translator: the input is already a
Vietnamese machine-converted draft. It is not intended for factual, legal,
medical, or other high-stakes text.

## Training data and procedure

The model was initialized from `DanVP/MoxhiMT-30-QT` and continued on
owner-curated Vietnamese draft→edited-text pairs. Training categories included
a filtered QT-convert pool (about 749k unique pairs), teacher-silver data
(about 49.6k pairs), a small teacher-gold seed (899 pairs), identity examples,
and about 150k synthetic name-swap pairs. Cumulative exposure reached
1,237,021,456 target labels. Raw training and holdout corpora are not
redistributed in this repository.

The published `model.safetensors` is a deterministic reverse mapping of the
training checkpoint into standard `MarianMTModel` tensor names. The release
gate verifies complete tensor coverage, exact expected key layout, successful
`AutoTokenizer`/`AutoModelForSeq2SeqLM` loading, and logits/argmax parity against
the training implementation. Exact hashes are recorded in
`release_manifest.json`.

## Limitations

- Training and validation focus on VietPhrase/QT drafts from Chinese web
  novels. News, technical, legal, medical, academic, business, conversational,
  and other genres are out of domain. The model may rewrite already-correct
  prose, corrupt specialist terminology, or alter the meaning; avoid using it
  as a general-purpose Vietnamese editor.
- Rare or out-of-distribution name inventories can cause large quality drops.
- Bare decoding may invent a digit, repeat text, omit content, or damage a
  proper name. The guarded application path reduces these failures but does
  not eliminate semantic errors.
- The checker still has a known blind spot for name-boundary resegmentation:
  individually attested words can be recombined into the wrong name span.
- Earlier checker revisions also conflated book-level name recap with
  invention and could flag legitimate repeated proper nouns; those two cases
  were corrected in checker v2.
- The model works one line at a time, so cross-sentence consistency depends on
  the input draft and surrounding application.

## License and attribution

The model repository is released under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Commercial use,
modification, and redistribution are permitted with attribution. Suggested
credit: **“vp2vi by DanVP”**, linking to
`https://huggingface.co/DanVP/vp2vi`.

The project owner authorized this public release on 2026-08-11. The model
repository contains model artifacts and documentation only; the separately
deployed browser runtime/Space has its own release boundary.