ELECTRA discriminator small
- pretrained with large Korean corpus datasets (30GB)
- 13.7M model parameters (followed google/electra-small-discriminator config)
- 32,000 vocab size
- trained for 1,000,000 steps
- build with lassl framework
pretrain-data
┣ korean_corpus.txt
┣ kowiki_latest.txt
┣ modu_dialogue_v1.2.txt
┣ modu_news_v1.1.txt
┣ modu_news_v2.0.txt
┣ modu_np_2021_v1.0.txt
┣ modu_np_v1.1.txt
┣ modu_spoken_v1.2.txt
┗ modu_written_v1.0.txt