Live triage console
Real posts from four disasters the model never saw in training, triaged by the fine-tuned model and ranked by urgency, the way a response desk would work them.
Priority queue
Category mix
Analyze a post
Paste any disaster-time post. The fine-tuned 2B model returns a strict JSON triage record. It runs on a CPU, so expect a few seconds.
Evaluation
Gold human labels from HumAID, on four disasters held out of training by event, not by random split. No test post, near-duplicate or event was used for training or model selection.
Category macro-F1 · 95% bootstrap CI
Recall on life-critical classes
Per-disaster macro-F1
Per-class F1
Structured fields vs teacher (silver) labels
Category is scored against human gold labels. Actionable / urgency / location / needs are scored against silver labels from a teacher LLM (gemma-4-31b-it) and are reported separately.
AutoScientist iterations
Win-rate is AutoScientist's judge score of each iteration against that run's own base model, so it is not comparable across model sizes.
All systems
Model
How CrisisLens was built, trained and evaluated, and where it should not be trusted.