KoELECTRA

Why we included this project

KoELECTRA is a Korean ELECTRA model that ships pretrained in base and small sizes, trained on 34GB of Korean text. The checkpoints are hosted on Hugging Face, so you can load them into the Transformers library with a single from_pretrained call in PyTorch or TensorFlow instead of pretraining a model yourself. The repo documents the pretraining setup, including TPU usage and hyperparameters, which makes reproducing or extending the work straightforward. It also includes fine-tuning results on standard Korean benchmarks, a useful reference when you are deciding between the v1 and v2 releases for something like sentiment analysis or NER. This is a focused model release rather than a general-purpose framework, so expect to fine-tune it for your own downstream task.