Calibrate Before Use:
          Improving Few-Shot Performance of Language Models

           Tony Z. Zhao * 1 Eric Wallace * 1 Shi Feng 2 Dan Klein 1 Sameer Singh 3

            Abstract                    Input: Subpar acting. Sentiment: Negative
                                    Input: Beautiful film. Sentiment: Positive
   GPT-3 can perform numerous tasks when pro-            Input: Amazing.     Sentiment:
   vided a natural language prompt that contains a      where the first two lines co ...