Russian Toxicity Classifier API

In the age of digital communication, the spread of harmful content has become a prevalent issue. To address this, the Russian Toxicity Classifier API aims to detect and classify potentially toxic comments in Russian language texts. By leveraging advanced machine learning techniques and models from the Hugging Face's transformers library, this service can identify whether a given text is neutral or toxic. This not only assists in content moderation but also provides a tool for researchers, developers, and platforms to better understand and manage online interactions in the Russian digital space.

Furthermore, this API is fully compatible with Label Studio and can be seamlessly integrated as a machine learning backend, enabling users to take advantage of the powerful labeling tools provided by Label Studio while benefiting from the toxicity classification capabilities of this service.

This project serves a simple Flask API that uses a model from the Hugging Face's transformers library to classify Russian text as either neutral or toxic. The model in use is s-nlp/russian_toxicity_classifier.

More details here: https://dzen.ru/a/ZOIIrENDVgjvKrGB (on russian)

Requirements

• Docker
• Docker Compose
• NVIDIA GPU (For CUDA operations)

Additional Information

The Docker container is set up to use NVIDIA GPUs, optimized for CUDA operations. Ensure that you have the necessary NVIDIA drivers installed on your machine.

Settings / Labeling Interface

For those using Label Studio, the following configuration can be set up for the labeling interface:

<View>
    <Text name="text" value="$text"/>
    <Choices name="sentiment" toName="text" showInLine="true">
        <Choice value="toxic" background="red"/>
        <Choice value="neutral" background="gray"/>
    </Choices>
</View>

This configuration allows users to label texts with either "toxic" or "neutral" sentiments. The choices are displayed in-line, with distinct colors for each sentiment.

Dependencies

• Flask (v1.1.2)
• transformers (v4.31.0)
• PyTorch (v2.0.1)
• gunicorn (v20.1.0)
• label-studio-sdk (v0.0.30)
• label-studio-ml (v1.0.9)
• rq (v1.15.1)

Setup and Installation

Copy compose file:

cp docker-compose.dist.yml docker-compose.yml

Building the Docker Image:

docker-compose build

Running the Application:

docker-compose up -d

API Endpoints

Health Check

You can check the health status of the service by simply accessing its root (/) URL.

• URL: /
• Method: GET

If everything is running correctly, you should see:

{
  "model_dir": null,
  "status": "UP",
  "v2": false
}

This indicates that the service is up and running.

Predict Toxicity

The Predict Toxicity endpoint allows users to submit an array of text strings in Russian. The service then evaluates each text string and returns predictions indicating whether each text is neutral or toxic. Additionally, a score representing the confidence of the prediction is provided.

• URL: /predict
• Method: POST

Data Params:

{
  "tasks": [
    {"data": {"text": "Your text 1"}},
    {"data": {"text": "Your text 2"}}
  ]
}

cURL Example:

curl -X POST "http://localhost:5000/predict" -H "Content-Type: application/json" -d '{
  "tasks": [
    {"data": {"text": "Your text 1"}},
    {"data": {"text": "Your text 2"}}
  ]
}'

Success Response:

[
  {
    "result": [
      {
        "from_name": "sentiment",
        "to_name": "text",
        "type": "choices",
        "value": {
          "choices": [
            "neutral"
          ]
        }
      }
    ],
    "score": 0.1234
  },
  ...
]

Links

• https://dzen.ru/a/ZOIIrENDVgjvKrGB
• https://huggingface.co/s-nlp/russian_toxicity_classifier
• https://github.com/HumanSignal/label-studio-ml-backend/tree/master