Japanese Sexual Moderation V2

🚀 Japanese Sexual Moderation V2

A fine - tuned model based on [studio - ousia/luke - japanese - large - lite](https://huggingface.co/studio - ousia/luke - japanese - large - lite) that scores whether short texts are sexual.

This model is trained using regression, and the output scores generally range from 0 to 1, but negative values or values exceeding 1 may occur. The model has not been trained on long texts, and it is assumed that the input will be split by line breaks.

• 0.0 - 0.2: Not sexual at all
• 0.2 - 0.4: Hardly contains sexual content
• 0.4 - 0.6: May contain sexual content
• 0.6 - 0.8: Contains sexual content
• 0.8 - 1.0: Highly sexual content

🚀 Quick Start

This model can quickly score whether short texts are sexual.

💻 Usage Examples

Basic Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "oshizo/japanese-sexual-moderation-v2"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(
    model_id,
    problem_type="regression"
)

texts = [
    "富士山は日本で一番高い山です。",
    "こんなになるまで我慢してたんだ...大変だったね",
    "俺は倒れる京子に近づき、担ぎ上げようと太ももの下に腕を通す。",
]

with torch.no_grad():
    encoding = tokenizer(texts, padding=True, max_length=64, return_tensors="pt")
    scores = model(**encoding).logits

# tensor([[-0.0045],
#         [ 0.2954],
#         [ 0.4391]])

📄 License

This project is licensed under the MIT license.