bert-disaster-tweets

bert-base-uncased fine-tuned on Kaggle Natural Language Processing with Disaster Tweets to classify whether a tweet is about a real disaster (disaster, label 1) or not (not_disaster, label 0).

Training: 6,090 tweets (stratified 80/20 split of the Kaggle train set), 3 epochs, lr 2e-5, batch size 32, max length 128, linear schedule with 10% warmup.

Files

• model.safetensors: PyTorch weights (stored in fp16), loadable with AutoModelForSequenceClassification.
• onnx/model.onnx: fp32 export; predictions identical to the PyTorch model on the validation set.
• onnx/model_quantized.onnx: per-channel dynamic int8 quantization; same validation accuracy (0.8424), but about 10% of individual predictions differ from the fp32 model.

Usage (transformers)

from transformers import pipeline
classifier = pipeline("text-classification", model="XiaoxiaoMa001/bert-disaster-tweets")
classifier("Forest fire near La Ronge Sask. Canada")

Usage (transformers.js)

import { pipeline } from "@huggingface/transformers";
const classifier = await pipeline("text-classification", "XiaoxiaoMa001/bert-disaster-tweets", { dtype: "fp32" });
await classifier("Forest fire near La Ronge Sask. Canada", { top_k: null });