NERDD

NER pipeline for the Disque Denúncia context, organized into training, calibration, and pseudolabelling subpipelines.

Current Structure

• src/: main source code.
• src/base_model_training/: base training and evaluation with nested CV.
• src/pseudolabelling/: pseudolabel generation, score-based split, and refit.
• src/calibration/: fit/apply reusable probability calibrators for the base model scores.
• src/tools/: auxiliary utilities.
• docs/: operational and architectural documentation.
• data/: training, test, and calibration datasets.

Prerequisites

• Git
• Python 3.11+
• pip

Quick Setup

git clone https://github.com/MLRG-CEFET-RJ/nerdd.git
cd nerdd
cd src
pip install -r requirements.txt

Next Steps

• Detailed installation: docs/INSTALL.md
• Runbook: docs/RUNBOOK.md
• Pipeline overview: docs/PIPELINE_OVERVIEW.md
• Architecture: docs/ARCHITECTURE.md
• Architectural decisions: docs/ARCHITECTURAL_DECISIONS.md
• Migration: docs/MIGRATION.md

Canonical Flow

1. Train the base model in src/base_model_training/.
2. Build a labeled calibration subset and fit a reusable calibrator artifact in src/calibration/.
3. Run large-corpus prediction in src/pseudolabelling/, optionally applying the calibrator during inference.

Contributing

Open an issue or PR with fixes and improvements.