This educational model fine-tunes distilbert/distilbert-base-uncased to
classify English tweets as DISASTER or NOT_DISASTER.
Validation results
Metric
Value
Accuracy
0.8516
Precision (DISASTER)
0.8754
Recall (DISASTER)
0.7630
F1 (DISASTER)
0.8154
The validation split contains 1,523 examples. The best checkpoint was selected
by validation F1 after the first training epoch.
Intended use and limitations
This model is a course project and demonstration. It must not be used as a
real emergency detection or public-safety system. The source dataset contains
ambiguous labels, exact duplicate texts, and informal language. The model can
also mistake simulations, false alarms, metaphors, or staged events for real
disasters.
