Statistics

The mean perplexity of the model of the greedy decoding is 2.723, while an untrained model's mean perplexity of the 390 examples leads 11.75.

We use the Rogue score 1 , 2 , L and sum of the ChatGPT with the same instruction for the Prompt Generator models and compare the rogue score. The high rogue score demonstrates the model will closely match the reference text for predictions. However, these metrics ignore that ChatGPT does the same instruction very differently from the reference text.