SKTBrain/KoBERT

Pretrained on 5M Korean Wikipedia sentences with SentencePiece tokenization, achieving a compact 8,002-token vocabulary (92M parameters vs. 110M for multilingual BERT). Supports PyTorch, ONNX, and MXNet-Gluon frameworks with ready-to-use model loading APIs. Demonstrates superior Korean NLP performance on sentiment analysis and named entity recognition tasks compared to Google's multilingual BERT baseline.

1,407 stars. No commits in the last 6 months.