A compact, fine-tuned BERT model to classify tweets as Disaster or Not Disaster. It is a lightweight disaster-intent classifier optimized for short-form social media text. Designed for fast inference, containerized deployment, and reproducible research.

IMPORTANT NOTE: Class imbalance was INTENTIONALLY preserved to reflect real-world disaster distribution. See training notebook for detailed rationale.

Model: bert-base-uncased
Author: sakibalfahim
Uploaded: 2025-12-22

Label mapping

0 → Not Disaster
1 → Disaster

How to use (example)

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo_id = 'sakibalfahim/CrisisPulse'
token = 'hf_xxx'  # use secure token or Hugging Face login
tokenizer = AutoTokenizer.from_pretrained(repo_id, token=token)
model = AutoModelForSequenceClassification.from_pretrained(repo_id, token=token)
model.eval()

inputs = tokenizer('Massive flood reported in downtown area', return_tensors='pt', truncation=True)
with torch.no_grad():
    logits = model(**inputs).logits
    pred = int(logits.argmax(-1)[0].item())
print('Prediction:', {0: 'Not Disaster', 1: 'Disaster'}[pred])

Demo

URL: https://huggingface.co/spaces/sakibalfahim/CrisisPulse

Training summary

Base model: bert-base-uncased
Training environment: NVIDIA GPU (PyTorch, Transformers)
Saved artifacts: Uploaded to this repository.

Intended use & limitations

Intended for research/demo use. Validate on your domain before any high-stakes use. Be cautious with domain shift, sarcasm, or non-English text.

Reproducibility

Check the notebook for exact preprocessing, hyperparameters, and seed.

License

Apache-2.0 License.

Contact

Author: sakibalfahim — via Hugging Face profile.