IndoBERT Base P1

IndoBERT Base (phase1 - uncased) is an embedding model that generates contextual representations of Indonesian text. It is a BERT-base architecture (124.5M parameters) pretrained on the Indo4B dataset (23.43 GB of text collected from social media, blogs, news, and websites) using masked language modeling and next sentence prediction objectives. The model is released under the MIT license and supports PyTorch, TensorFlow, and JAX frameworks. As part of the IndoNLU benchmark, this model was evaluated across twelve tasks ranging from single-sentence classification to pair-sentence sequence labeling, covering diverse domains and complexity levels. The model serves as a baseline for the IndoNLU benchmark suite, which was introduced at AACL-IJCNLP 2020. The Indo4B training corpus was curated from publicly available Indonesian sources including social media, blogs, news, and websites to ensure broad linguistic coverage. The model has been widely adopted within the Indonesian NLP ecosystem, with over 855,000 monthly downloads, 100 Spaces, and 139 finetuned variants. It is tagged for Feature Extraction as its primary task and is available under the MIT license. The official indobenchmark-toolkit repository provides evaluation and benchmarking utilities for this model.