ClearSignal — Production Content Moderation API

Project Status

[!IMPORTANT] Current Status: 🟢 Phase 1 Complete (Data Pipeline) | 🔴 Phases 2–5 Planned (In Development)

• Implemented: Phase 1 Data Pipeline (scripts/prepare_data.py, scripts/auto_label.py, scripts/clean_data.py).
• NOT Yet Implemented:
  • Phase 2: Model Fine-Tuning (DeBERTa-v3 multi-label classification)
  • Phase 3: FastAPI Backend & SHAP Token Attribution
  • Phase 4: MLOps Containerization & AWS EC2/ECR Deployment
  • Phase 5: MMD Input Drift Monitoring & CUPED A/B Testing Infrastructure

Overview

ClearSignal is designed as a multi-label content moderation system that categorizes text across 8 harmful content dimensions (toxic, hate_speech, harassment, misinformation, spam, self_harm, sexual_content, clean) with confidence scores and explainability features.

Currently, Phase 1 (Data Preparation & Labelling Pipeline) is fully implemented. The data pipeline aggregates public datasets, runs automated labelling with a pre-trained transformer model, and produces a cleaned dataset ready for model training.

Implemented Data Pipeline (Phase 1)

The scripts/ directory contains the working data preparation suite:

1. Data Ingestion (scripts/prepare_data.py):

   • Downloads 1,000 tweets from tweets_hate_speech_detection (500 hate, 500 clean).
   • Downloads 1,000 comments from ucberkeley-dlab/measuring-hate-speech (500 hateful, 500 clean).
   • Pre-Sampling Deduplication & Determinism: Fixed a data leakage bug where (~3.4 annotator entries per text) caused inflated duplicate counts downstream. Deduplication by text is performed to shuffling () and sampling, ensuring 100% deterministic and unique post selection.