chiefj/toxicity-worker repository overview

⁠Stateless Toxicity Worker

Production-ready multi-label toxicity classification worker service built with Hexagonal Architecture (Ports and Adapters) using unitary/toxic-bert run with ONNX Runtime on CPU.

⁠Installation and Dependencies

The service uses:

• fastapi & uvicorn for the REST API
• optimum[onnxruntime] for exporting and running the model on ONNX Runtime CPU
• opentelemetry-api & opentelemetry-sdk for tracing

⁠Configuration & Environment Variables

⁠Docker Deployment

⁠Building the Image

docker build -t chiefj/toxicity-worker:latest -f build/Dockerfile .

⁠Running with Docker

Run the container by passing the environment variables:

docker run -d \
  -p 8008:8008 \
  -e PORT=8008 \
  -e TOXICITY_MODEL_ID=unitary/toxic-bert \
  --name toxicity-worker \
  chiefj/toxicity-worker:latest

⁠API Endpoint Contract

Defined in contracts/openapi/v1.yaml.

⁠POST /healthz / GET /healthz

Response:

{
  "status": "ok",
  "model": "unitary/toxic-bert",
  "model_id": "unitary/toxic-bert"
}

⁠POST /score

Request:

{
  "text": "response text here"
}

Response:

{
  "toxicity": 0.02,
  "severe_toxicity": 0.00,
  "obscene": 0.01,
  "threat": 0.00,
  "insult": 0.02,
  "identity_hate": 0.00
}

For texts exceeding 510 tokens, the response will also include the long response strategy:

{
  "toxicity": 0.02,
  "severe_toxicity": 0.00,
  "obscene": 0.01,
  "threat": 0.00,
  "insult": 0.02,
  "identity_hate": 0.00,
  "long_response_strategy": "max_of_two_passes"
}

⁠Running Tests

Run the test suite using pytest with the following command:

cd packages/python/toxicity-worker
python3 -m pytest tests/ -v --tb=short --cov=src --cov-report=term-missing