|
--- |
|
license: bigscience-bloom-rail-1.0 |
|
datasets: |
|
- jigsaw_toxicity_pred |
|
language: |
|
- en |
|
metrics: |
|
- accuracy |
|
- f1 |
|
library_name: transformers |
|
pipeline_tag: text-classification |
|
--- |
|
|
|
## Model description |
|
This model is a fine-tuned version of the [bert-base-uncased](https://huggingface.co/transformers/model_doc/bert.html) model to classify toxic comments. \ |
|
The BERT model is finetuned using adversarial training to boost robustness against textual adversarial attacks. |
|
|
|
## How to use |
|
|
|
You can use the model with the following code. |
|
|
|
```python |
|
from transformers import BertForSequenceClassification, BertTokenizer, TextClassificationPipeline |
|
model_path = "JiaqiLee/robust-bert-jigsaw" |
|
tokenizer = BertTokenizer.from_pretrained(model_path) |
|
model = BertForSequenceClassification.from_pretrained(model_path, num_labels=2) |
|
pipeline = TextClassificationPipeline(model=model, tokenizer=tokenizer) |
|
print(pipeline("You're a fucking nerd.")) |
|
``` |
|
|
|
## Training data |
|
The training data comes from this [Kaggle competition](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data). We use 90% of the `train.csv` data to train the model. \ |
|
We augment original training data with adversarial examples generated by PWWS, TextBugger and TextFooler. |
|
|
|
## Evaluation results |
|
|
|
The model achieves 0.95 AUC in a 1500 rows held-out test set. |