---
language:
- en
license: apache-2.0
base_model: bert-base-uncased
pipeline_tag: text-classification
tags:
- text-classification
- bert
- disaster-tweets
- kaggle
datasets:
- startificial/nlp-getting-started
metrics:
- accuracy
- f1
---

# Disaster Tweet Detection（灾难推文检测）

基于 `bert-base-uncased` 微调的英文推文二分类模型，判断一条推文是否**关于真实灾难**
（Kaggle 竞赛 *Real or Not? NLP with Disaster Tweets*，`nlp-getting-started`）。

## 训练数据

- 来源：Kaggle 灾难推文数据集（7613 条训练样本）
- 标签：0 = 非灾难推文，1 = 灾难推文
- 训练：2 个 epoch，lr=2e-5，batch=16，max_len=128

## 验证集指标

| 指标 | 数值 |
| --- | --- |
| Accuracy | **0.8451** |
| F1 (binary) | **0.8168** |
| Precision | 0.8297 |
| Recall | 0.8043 |

## 使用方式

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "Operaer/disaster-tweet-bert"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

text = "Forest fire near La Ronge Sask. Canada"
enc = tokenizer(text, truncation=True, padding=True,
                max_length=128, return_tensors="pt")
with torch.no_grad():
    logits = model(**enc).logits
pred = int(torch.argmax(logits, dim=1))
print("灾难推文" if pred == 1 else "非灾难推文")
```

## 说明

- 本模型为《自然语言处理》课程作业产物；
- 微调代码：`train.py`；交互式 Demo：`app.py` / Gradio Space。