Political Misinformation Classifier
Applied ML: BERT fine-tuning, evaluation discipline, honest limitations
- Academic
- 2025
- Python BERT NLP Machine learning Model evaluation Data pipelines
- Private repository
ML work with an engineering mindset: the artifact is not only weights, it is the preprocessing contract, evaluation notebook, and clear statement of where the model will lie.
Problem
Keyword filters miss nuanced propaganda. The project tested whether a fine-tuned transformer could add signal while staying humble about politics, language shift, and dataset bias.
Model pipeline
- Text preprocessing aligned to tokenizer and task (no silent truncation surprises).
- Fine-tuned BERT classifier for political misinformation labels.
- Versioned training scripts and frozen splits for reproducible comparison.
Evaluation
Reported precision/recall tradeoffs, confusion patterns, and qualitative failures, not a single leaderboard number. The goal is to know when the model should refuse to classify.
Limitations
Temporal and demographic drift break political classifiers silently. These models belong behind human review, monitoring, and governance, not as autonomous truth engines.
What I learned
Most of the “backend” in ML is evaluation and data contracts; the forward pass is the short part.