distilbert-base-sexism-detector
This is a fine-tuned model of distilbert-base on the Explainable Detection of Online Sexism (EDOS) dataset. It is intended to be used as a classification model for identifying tweets (0 - not sexist; 1 - sexist).
This is a light model with an 81.2 F1 score. Use this model for fase prediction using the online API, if you like to see our best model with 86.3 F1 score , use this link.
Classification examples (use these example in the Hosted Inference API in the right panel ):
| Prediction | Tweet |
|---|---|
| sexist | Every woman wants to be a model. It's codeword for "I get everything for free and people want me" |
| not sexist | basically I placed more value on her than I should then? |
More Details
For more details about the datasets and eval results, see (we will updated the page with our paper link)
How to use
from transformers import AutoModelForSequenceClassification, AutoTokenizer,pipeline
import torch
model = AutoModelForSequenceClassification.from_pretrained('NLP-LTU/distilbert-sexism-detector')
tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased')
classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
prediction=classifier("Every woman wants to be a model. It's codeword for 'I get everything for free and people want me' ")
label_pred = 'not sexist' if prediction == 0 else 'sexist'
print(label_pred)
precision recall f1-score support
not sexsit 0.9000 0.9264 0.9130 3030
sexist 0.7469 0.6784 0.7110 970
accuracy 0.8662 4000
macro avg 0.8234 0.8024 0.8120 4000
weighted avg 0.8628 0.8662 0.8640 4000