KoBERT

Why we included this project

Multilingual BERT models often lose accuracy on Korean, and KoBERT exists to fix exactly that: a cased BERT trained on 5 million sentences from the Korean Wikipedia. The vocabulary is a compact 8,002-token SentencePiece set, which keeps the model at 92M parameters versus the multilingual baseline's 110M. You load the weights through a simple Python package, and alongside the usual PyTorch weights there are ONNX and MXNet-Gluon versions, handy if you serve with ONNX Runtime. The repo also bundles fine-tuning examples for Naver sentiment analysis and a BERT-CRF named entity recognizer, giving classification and sequence-labeling work ready-made starting points. For teams that would rather fine-tune a Korean-native BERT than wrestle with a generic multilingual checkpoint, KoBERT is a solid place to start.