Deberta V2 Base Japanese Finetuned QAe

🚀 DeBERTa-V2 Japanese Fine-tuned for QA

This model is fine-tuned from deberta-v2-base-japanese for Question-Answering tasks, offering high accuracy for QA in the driving domain and general scenarios.

🚀 Quick Start

First, install the necessary libraries: transformers, pytorch, sentencepiece, and Juman++. Then, you can execute the following code to perform a Question-Answering task:

import torch
from transformers import AutoTokenizer, AutoModelForQuestionAnswering

tokenizer = AutoTokenizer.from_pretrained('ku-nlp/deberta-v2-base-japanese')
model=AutoModelForQuestionAnswering.from_pretrained('Mizuiro-sakura/deberta-v2-base-japanese-finetuned-QAe') # Load the pre-trained model

text={
    'context':'私の名前はEIMIです。好きな食べ物は苺です。 趣味は皆さんと会話することです。',
    'question' :'好きな食べ物は何ですか'
}

input_ids=tokenizer.encode(text['question'],text['context']) # Perform morphological analysis and convert to codes using the tokenizer
output= model(torch.tensor([input_ids])) # Analyze using the pre-trained model
prediction = tokenizer.decode(input_ids[torch.argmax(output.start_logits): torch.argmax(output.end_logits)]) # Extract the part corresponding to the answer
print(prediction)

✨ Features

• Fine-tuned for QA: Based on deberta-v2-base-japanese, fine-tuned using the Driving Domain QA Dataset (DDQA) for better performance in Question-Answering tasks.
• Multilingual Data Training: Trained on Japanese Wikipedia, CC-100, and OSCAR, enabling it to handle a wide range of Japanese texts.

📦 Installation

You need to install transformers, pytorch, sentencepiece, and Juman++ to use this model.

📚 Documentation

Model Details

This model is a fine-tuned version of deberta-v2-base-japanese for Question-Answering tasks. It is trained on the DDQA dataset, which can be used for general Question-Answering tasks, especially in the driving domain.

Model Accuracy

• Exact Match: 0.8038277511961722
• F1 Score: 0.8959389668095072

What is deberta-v2-base-japanese?

It is a model trained on Japanese Wikipedia (3.2GB), CC-100 (85GB), and OSCAR (54GB). It was released by the Kurohashi Lab at Kyoto University.

Acknowledgments

I would like to thank the Kurohashi Lab at Kyoto University for releasing the deberta-v2-base-japanese model.

📄 License

This project is licensed under the MIT License.

Tags

• pytorch
• deberta
• deberta-v2
• question-answering
• squad

Datasets

• wikipedia
• cc100
• oscar

Metrics

• accuracy