Skip to Main Content

Prompt tuning with only 20K trainable parameters (vs 11B model parameters) matched full fine-tuning on SuperGLUE at scal.Lester et al., 'The Power of Scale for Parameter-E…