Add >20B trained ckpt

Update model with >10B trained ckpt, change model type to bfloat16(match pretrain)

update model w/ modified tokenizer, trained >5B tokens

Add 5B korean trained model, test version.