Fireclow/indobert-sentiment-toxicity-multitask

IndoBERT Multitask — Sentiment & Toxicity (Indonesian)

Fine-tuned indobenchmark/indobert-base-p1 for Indonesian social media comments:

Task	Labels
Sentiment	negative, neutral, positive
Toxicity	non_toxic, toxic

Files

• best_model.pt — multitask checkpoint (encoder + two classification heads)
• label_mappings.pt — id ↔ label maps
• tokenizer/ — tokenizer files

Usage with this project

Set in Streamlit secrets or environment:

HF_MODEL_REPO=Fireclow/indobert-sentiment-toxicity-multitask

Test metrics (held-out split)

• Sentiment accuracy: ~95%
• Toxicity accuracy: ~92%

University NLP final project.