Fake News Classifier - Riski Linardi Portfolio
A BERT model fine-tuned on the PolitiFact fact-check dataset. Paste a statement and the model returns a verdict, confidence, and optionally the tokens that drove the decision.
Try an example
About this model
The classifier is bert-base-uncased fine-tuned for three epochs on roughly 21,000 statements from PolitiFact, with labels collapsed from six fact-check categories into a binary truth-versus-fake target. Training notebook, comparison against classical baselines, and the FastAPI service code are in the repository.
Token attributions use integrated gradients via transformers-interpret. Raw attention weights are deliberately not used because they are not a reliable explanation signal for classification.
This is a portfolio project. The model reflects the biases and limitations of its training data and is not a substitute for human fact-checking.