CrisisConnect crisis priority classifier

Classifies a crisis-related text as Low, Medium, High or Critical information priority, following the TREC Incident Streams definition: how urgently an emergency manager needs to see it. Fine-tuned from crisistransformers/CT-M1-Complete on 35571 TREC-IS tweets.

On TREC-IS events held out from training (10210 tweets): accuracy 0.474, macro F1 0.416, off by two or more levels 0.102. For "needs prompt attention" (High or Critical): precision 0.396, recall 0.602.

Priority is subjective and TREC-IS annotations are noisy, so treat outputs as triage hints for a human reviewer.

Trained with train_crisis_models.ipynb, seed 42. Full metrics: metrics.json.