IndoNLP/indonlu

Encompasses 12 diverse NLU tasks spanning sequence classification, tagging, and semantic similarity, with evaluation infrastructure on CodaLab for reproducible benchmarking. The IndoBERT models employ a two-phase pretraining approach on Indo4B—a 4-billion-word, 20+ GB Indonesian corpus—with both base and lite variants available on Hugging Face. Also includes FastText embeddings trained on the same corpus, enabling practitioners to choose between transformer-based and lightweight word vector approaches for downstream Indonesian NLP tasks.

638 stars. No commits in the last 6 months.