trained using OSCAR dataset

vocab size 50000