IndoBERTweet-IdentityAttack

Model Description

IndoBERTweet fine-tuned on IndoToxic2024 dataset, with an accuracy of 0.89 and macro-F1 of 0.78. Performances are obtained through stratified 10-fold cross-validation.

Supported Tokenizer

• indolem/indobertweet-base-uncased

Example Code

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
# Specify the model and tokenizer name
model_name = "Exqrch/IndoBERTweet-IdentityAttack"
tokenizer_name = "indolem/indobertweet-base-uncased"
# Load the pre-trained model
model = AutoModelForSequenceClassification.from_pretrained(model_name)
# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
text = "selamat pagi semua!"
output = model(**tokenizer(text, return_tensors="pt"))
logits = output.logits
# Get the predicted class label
predicted_class = torch.argmax(logits, dim=-1).item()
print(predicted_class)
--- Output ---
> 0
--- End of Output ---

Limitations

Trained only on Indonesian texts. No information on code-switched text performance.

Sample Output

Model name: Exqrch/IndoBERTweet-IdentityAttack
Text 1: ayolah, jaga kebersihan bersama
Prediction: 0
Text 2: dia itu loh, udah hitam, dengkil lagi
Prediction: 1

Citation

If used, please cite:

@article{susanto2024indotoxic2024,
      title={IndoToxic2024: A Demographically-Enriched Dataset of Hate Speech and Toxicity Types for Indonesian Language}, 
      author={Lucky Susanto and Musa Izzanardi Wijanarko and Prasetia Anugrah Pratama and Traci Hong and Ika Idris and Alham Fikri Aji and Derry Wijaya},
      year={2024},
      eprint={2406.19349},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.19349}, 
}