Roberta Base Japanese Jsnli

🚀 roberta-base-japanese-jsnli

This model is a fine-tuned version of nlp-waseda/roberta-base-japanese on the JSNLI dataset. It is designed for text classification tasks, especially natural language inference. The model achieves high accuracy on the evaluation set, making it a reliable choice for related NLP applications.

🚀 Quick Start

Prerequisites

The input text should be segmented into words by Juman++ in advance.

✨ Features

• Zero-shot Classification: Capable of performing zero-shot classification tasks, allowing users to classify text into predefined labels without specific training on those labels.
• High Accuracy: Achieves an accuracy of 0.9328 on the evaluation set, demonstrating its effectiveness in natural language inference tasks.

📦 Installation

This section does not contain specific installation steps, so it is skipped.

💻 Usage Examples

Basic Usage

from transformers import pipeline
from pyknp import Juman

juman = Juman()

classifier = pipeline("zero-shot-classification", model="Formzu/roberta-base-japanese-jsnli")

sequence_to_classify = " ".join([tok.midasi for tok in juman.analysis("いつか世界を見る。").mrph_list()])

candidate_labels = ['旅行', '料理', '踊り']
out = classifier(sequence_to_classify, candidate_labels, hypothesis_template="この 例 は {} です 。")
print(out)
#{'sequence': 'いつか 世界 を 見る 。', 
# 'labels': ['旅行', '踊り', '料理'], 
# 'scores': [0.8998081684112549, 0.06059670448303223, 0.03959512338042259]}

Advanced Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
from pyknp import Juman

juman = Juman()

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

model_name = "Formzu/roberta-base-japanese-jsnli"
model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name)

premise = " ".join([tok.midasi for tok in juman.analysis("いつか世界を見る。").mrph_list()])
label = '旅行'
hypothesis = f'この 例 は {label} です 。'

input = tokenizer.encode(premise, hypothesis, return_tensors='pt').to(device)
with torch.no_grad():
    logits = model(input)["logits"][0]
    probs = logits.softmax(dim=-1)
    print(probs.cpu().numpy(), logits.cpu().numpy())
#[0.82168734 0.1744363  0.00387629] [ 2.3362164   0.78641605 -3.0202653 ]

📚 Documentation

Model Information

Evaluation Results

This model achieves the following results on the evaluation set:

• Loss: 0.2039
• Accuracy: 0.9328

🔧 Technical Details

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3.0

Training results

Framework versions

• Transformers 4.21.2
• Pytorch 1.12.1+cu116
• Datasets 2.4.0
• Tokenizers 0.12.1

📄 License

This project is licensed under the CC BY-SA 4.0 license.