BERT Disaster Tweet Classifier
Fully fine-tuned on 5,987 cleaned disaster tweet training examples, with 748 validation and 749 held-out test examples. Seed 42. All encoder and classifier parameters updated for 3 epochs. Best checkpoint selected by validation F1.
Test accuracy: 0.82109479. Binary disaster-class F1: 0.78387097. Source data: Kaggle NLP Getting Started, downloaded from startificial/nlp-getting-started. This is a local held-out split, not the Kaggle leaderboard.
Labels: 0=non_disaster, 1=disaster. English only, max length 128 tokens. Not for emergency response decisions. Softmax scores are not calibrated event probabilities.
Demo and training source: https://huggingface.co/spaces/jp2023213625/disaster-tweet-bert-homework
The browser demo uses an INT8 ONNX derivative, with separately reported test F1 0.77248677.
- Downloads last month
- 32
Model tree for jp2023213625/bert-disaster-tweet-homework
Base model
google-bert/bert-base-uncased