Model Card for BERT hate offensive tweets

BERT base uncased trained on the data that can be found here: MartynaKopyta/hate_offensive_tweets to classify tweets as 0 - hate, 1 - offensive or 2 - neither.

You can find the notebook used for training in my GitHub repo: MartynaKopyta/BERT_FINE-TUNING.

Model Details

Finetuned from model bert-base-uncased

Bias, Risks, and Limitations

The dataset was not big enough for BERT to learn to classify 3 classes accurately, it is right 3/4 times.

How to Get Started with the Model

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained('MartynaKopyta/BERT_hate_offensive_tweets')
tokenizer = AutoTokenizer.from_pretrained('MartynaKopyta/BERT_hate_offensive_tweets')

Training Hyperparameters

batch size:16
learning rate:2e-5
epochs:3

Evaluation

Accuracy: 0.779373368146214

Classification Report:

              precision    recall  f1-score   support

           0       0.74      0.68      0.71      1532
           1       0.85      0.88      0.87      1532
           2       0.74      0.78      0.76      1532

    accuracy                           0.78      4596
   macro avg       0.78      0.78      0.78      4596
weighted avg       0.78      0.78      0.78      4596

Confusion Matrix:
[[1043   96  393]
 [ 169 1343   20]
 [ 204  132 1196]]

MCC: 0.670