indonlu

Why we included this project

For anyone building or evaluating Indonesian-language NLP systems, IndoNLU is the benchmark to know. It bundles 12 downstream tasks covering sentiment, named-entity recognition, question answering, and similar NLU work, each with train and validation sets plus a masked test set so results stay comparable to the published numbers. The repo also ships IndoBERT and IndoBERT-lite, pretrained on a roughly 4-billion-word Indonesian corpus, along with the Indo4B text data and FastText embeddings used to train them. The starter code makes it easy to load a model and reproduce the reported results, which is handy when you want a solid baseline before adapting something to your own pipeline. If you work on low-resource or non-English language modeling, this is a well-documented reference for both evaluation and further pretraining.