ViBidLEQA_large: A Vietnamese Bidding Law Extractive Question Answering Model

Overview

ViBidLEQA_large is an Extractive Question-Answering (EQA) model specifically developed for the Vietnamese bidding law domain. Built upon the nguyenvulebinh/vi-mrc-large architecture and fine-tuned with a specialized bidding law dataset, this model achieves state-of-the-art performance in extracting precise answers from legal documents for bidding law queries.

Model Description

• Task: Extractive Question Answering
• Domain: Vietnamese Bidding Law
• Base Model: nguyenvulebinh/vi-mrc-large
• Approach: Fine-tuning
• Language: Vietnamese

Dataset

The ViBidLQA dataset consists of:

• Training set: 5,300 samples
• Test set: 1,000 samples
• Data Creation Process:
  • Training data was automatically generated using Claude 3.5 Sonnet and validated by two legal experts
  • The test set was manually created and verified by two Vietnamese legal experts
  • All samples focus on Vietnamese bidding law content

Performance

Our model achieves exceptional performance on the test set:

Metric	Score
Exact Match	88.30
F1-Score	94.25

Usage

from transformers import AutoTokenizer, AutoModelForQuestionAnswering
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("ntphuc149/ViBidLEQA_large")
model = AutoModelForQuestionAnswering.from_pretrained("ntphuc149/ViBidLEQA_large")

# Example usage
question = "Thế nào là đấu thầu hạn chế?"
context = "Đấu thầu hạn chế là phương thức lựa chọn nhà thầu trong đó chỉ một số nhà thầu đáp ứng yêu cầu về năng lực và kinh nghiệm được bên mời thầu mời tham gia."

# Tokenize input
inputs = tokenizer(
    question,
    context,
    return_tensors="pt",
    max_length=512,
    truncation=True,
    padding=True
)

# Get model predictions
with torch.no_grad():
    outputs = model(**inputs)

# Get answer span
answer_start = torch.argmax(outputs.start_logits)
answer_end = torch.argmax(outputs.end_logits) + 1

answer = tokenizer.decode(inputs.input_ids[0][answer_start:answer_end])
print(f"Question: {question}")
print(f"Answer: {answer}")

Applications

This model is advantageous for:

• Legal document analysis systems
• Bidding law information retrieval systems
• Legal advisory chatbots
• Automated question-answering systems for bidding law
• Legal research and documentation tools

Limitations

• Domain Specificity: The model is specifically trained for Vietnamese bidding law and may not generalize well to other legal domains
• Language Constraint: Optimized for Vietnamese language only
• Context Length: Maximum input length is 512 tokens
• Legal Disclaimer: Should be used as a reference tool, not as a replacement for professional legal advice

Citation

comming soon...

Contact

For questions, feedback, or collaborations:

• Email: [email protected]
• GitHub Issues: @ntphuc149
• HuggingFace: @ntphuc149

License

This project is licensed under the MIT License - see the LICENSE file for details.