metadata

license: mit
language:
  - en
  - ja
  - zh
  - ko
metrics:
  - accuracy
base_model: google-bert/bert-base-multilingual-cased
pipeline_tag: text-classification
tags:
  - sex
  - filename
  - dectection
  - content
  - mbert
  - Multilingual

Model Card for Model ID

Detect sexual content in text or file names.

Model Details

Model Description

Developed by: liu wei
License: MIT
Finetuned from model: bert-base-multilingual-cased
Task: Simple Classification
Language: Multilingual
Max Length: 128
Updated Time: 2024-8-22

Model Training Information

Training Dataset Size: 100,000 manually annotated data with noise
Data Distribution: 50:50
Batch Size: 8
Epochs: 5
Accuracy: 92%
F1: 92%

Buy me a cup of coffee,thanks

Uses

Supports multiple languages, such as English, Chinese, Japanese, etc.
Use for detect content submitted by users in forums, magnetic search engines, cloud disks, etc.
Detect semantics and variant content, Porn movie numbers or variant file names.
Compared with GPT4O-mini, The detection accuracy is greatly improved.

Examples

Example English

predict("Tiffany Doll - Wine Makes Me Anal (31.03.2018)_1080p.mp4")

{
    "predictions": 1,
    "label": "Sexual"
}

Example Chinese

predict("橙子 · 保安和女业主的一夜春宵。路见不平拔刀相助，救下苏姐，以身相许！")

{
    "predictions": 1,
    "label": "Sexual"
}

Example Japanese

predict("MILK-217-UNCENSORED-LEAKピタコス Gカップ痴女 完全着衣で濃密5PLAY 椿りか 580 2.TS")

{
    "predictions": 1,
    "label": "Sexual"
}

Example Porn Movie Numbers

predict("DVAJ-548_CH_SD")

{
    "predictions": 1,
    "label": "Sexual"
}

How to Get Started with the Model

step 1:

Create a python file under this model, such as 'use_model.py'

import torch
from transformers import BertForSequenceClassification, BertTokenizer

# load model
tokenizer = BertTokenizer.from_pretrained("uget/sexual_content_dection")
model = BertForSequenceClassification.from_pretrained("uget/sexual_content_dection")

def predict(text):
    encoding = tokenizer(text, return_tensors="pt")
    encoding = {k: v.to(model.device) for k,v in encoding.items()}

    outputs = model(**encoding)
    probs = torch.sigmoid(outputs.logits)
    
    predictions = torch.argmax(probs, dim=-1)
    label_map = {0: "None", 1: "Sexual"}
    predicted_label = label_map[predictions.item()]
    print(f"Predictions:{predictions.item()}, Label:{predicted_label}")
    return {"predictions": predictions.item(), "label": predicted_label}

predict("Tiffany Doll - Wine Makes Me Anal (31.03.2018)_1080p.mp4")

step 2:

Run

python3 use_model.py

Response JSON

{
    "predictions": 1,
    "label": "Sexual"
}

Explanation

The results only include two situations:

predictions-0 Not Dectection sexual content;
predictions-1 Sexual content was detected.

Buy me a cup of coffee,thanks

Model Card Contact

Email: [email protected]