CollinMa/disaster-tweets-bert

BERT fine-tuned on Disaster Tweets

Full-parameter fine-tuning of English BERT for binary disaster-tweet classification. Labels: NOT_DISASTER (0), DISASTER (1). This model classifies text and does not verify events.

Held-out evaluation

Model	Accuracy	Precision	Recall	F1
TF-IDF + Logistic Regression	0.7864	0.8093	0.6520	0.7222
BERT full fine-tuning	0.8238	0.8258	0.7429	0.7822

749 test examples; duplicate removal before a seeded, stratified 80/10/10 split. Best checkpoint selected using validation F1 only. No Kaggle leaderboard score is claimed.

Training

3 epochs, learning rate 2e-05, batch size 8, gradient accumulation 2, max length 128, AdamW, dynamic padding, bf16. Seed 42. Base revision: 86b5e0934494bd15c9632b12f734a8a67f723594. Dataset mirror revision: 896f918ab4d22b0ad24f8e2d01b84acf0742c050. Original data source: https://www.kaggle.com/competitions/nlp-getting-started The source data's usage conditions continue to apply independently of this repository's model metadata.

Inference

from transformers import pipeline
classifier = pipeline("text-classification", model="CollinMa/disaster-tweets-bert")
classifier("Residents are evacuating as a wildfire approaches.", truncation=True, max_length=128)

Scores are uncalibrated softmax outputs. English input only; inputs longer than 128 tokens are truncated. Full training summary: training_summary.json.

Browser deployment

ONNX model: onnx/model_quantized.onnx. 浏览器实际使用同一检查点的 INT8 ONNX 版本，大小 182.2 MB。该版本在同一独立测试集（batch=1）上 Accuracy=0.8238，Precision=0.8327，Recall=0.7335，F1=0.7800。这次评估用于记录量化对实际部署模型的影响，不再用于选择训练检查点。