vit5-translation-vi2en-v1.1

Model Card for ViT5 Translation Model

A sequence-to-sequence translation model based on VietAI ViT5-base, fine-tuned for Vietnamese to English machine translation.
This model is intended for general-purpose translation tasks, both academic and production-oriented.

---

Model Details

Model Description

This model is an encoder–decoder Transformer designed for text-to-text generation tasks such as translation.
It is fine-tuned from VietAI/vit5-base, a pretrained multilingual T5 variant optimized for Vietnamese NLP tasks.

• Developed by: tnguyen20604
• Funded by [optional]: N/A
• Shared by: tnguyen20604
• Model type: Encoder–Decoder (Text-to-Text Transformer)
• Language(s): Vietnamese, English
• License: apache-2.0
• Fine-tuned from: VietAI/vit5-base

---

Model Sources

• Repository: https://huggingface.co/tnguyen20604/vit5-translation-vi2en-v1.1
• Base Model: https://huggingface.co/VietAI/vit5-base
• Dataset: https://huggingface.co/datasets/harouzie/vi_en-translation
• Synthetic Dataset: https://www.kaggle.com/datasets/nguyentran20604/synthentic-data

---

Synthetic Data

In addition to real bilingual corpora, this model was also trained using a synthetic English–Vietnamese dataset created to improve translation robustness and domain diversity.

Synthetic Dataset

• Source: https://www.kaggle.com/datasets/nguyentran20604/synthentic-data
• Contains machine-generated Vietnamese to English parallel sentences.
• Designed to supplement low-resource Vietnamese MT tasks and expand linguistic diversity.

Synthetic Data Generation Process

Synthetic data was generated using a large language model (LLM) through a multi-step workflow:

1. Prompting a large LLM (e.g., GPT or Qwen) to generate bilingual sentence pairs with:

   • correct semantic alignment
   • natural fluency in both languages
   • diverse sentence structures
   • a variety of writing styles (formal, casual, conversational, instructional)

2. Filtering and Cleaning

   • removed low-quality or incomplete generations
   • removed hallucinated or irrelevant translations
   • used automatic scoring (LLM-based + heuristic rules) to ensure source–target consistency
   • deduplicated repeated patterns

3. Formatting

   • normalized Unicode
   • standardized punctuation
   • removed extremely short or extremely long sentences
   • converted examples into a clean parallel translation format compatible with Seq2Seq training

Purpose of Synthetic Data

The synthetic dataset was introduced to:

• improve generalization on unseen sentence structures
• increase translation robustness across multiple domains
• expand vocabulary coverage and idiomatic expressions
• reduce overfitting to the original dataset

---

Uses

Direct Use

• Machine translation Viet → Eng
• Text rewriting (via Seq2Seq generation)
• Academic NLP experiments
• MT benchmarking

Downstream Use

• Fine-tuning thêm cho domain-specific translation (y khoa, pháp lý, kỹ thuật)
• Dùng trong pipeline dịch tự động hoặc chatbot đa ngữ

Out-of-Scope Use

• Not suitable for evaluating legal or medical content.
• Does not guarantee full accuracy for texts containing highly specialized terminology.
• Not appropriate for processing sensitive data or personally identifiable information (PII).

---

Bias, Risks, and Limitations

• The training data originates from open-source corpora, which may introduce stylistic or domain bias.
• The model may produce incorrect translations in cases such as:
  • ambiguous sentences
  • culturally specific expressions
  • very long or structurally complex sentences
• Some translations may lose nuance, tone, or contextual meaning.

Recommendations

Users should manually review the translations when used in professional, safety-critical, or high-importance scenarios.

---

How to Get Started

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "tnguyen20604/vit5-translation-vi2en-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

text = "Tôi yêu học máy."
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs)

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