ALT Amazon product page where the title of a strange product is "sorry but I can't provide the analysis you're looking for"
ALT Screenshot of Amazon product page for a pull-out kitchen storage basket, with the text "I'm Sorry but 2nd Floor Pull Blue White is not Clear Enough to Suggest Related Product Keywords. Can You Please Provide More Context or Information About What You are Looking for" where the product title would normally be written.
ALT Screenshot of Amazon product page for a jewelry store display, with the text "Sorry there seems to be an error in the provided prompt there is no specific product or keyword given. Could you please provide more information or specify a particular product or category-size1" where the product title would normally be written.
ALT Screenshot of Amazon product page for an electronic handheld device, with the text "Apologies but I'm unable to fulfill that request-size1" where the product title would normally be written.
ALT Screenshot of Amazon product page for an outdoor patio table, with the text "Sorry but I can't generate a response without changing the input-Green" where the product title would normally be written.
ALT GPT-3’s accuracy depends highly on prompt format. We next keep the set of training examples and permutations fixed but vary the prompt format. We focus on SST-2, and we manually design an additional 14 prompt formats. The formats include question-answer templates, conversation- style templates, prompts that resemble Web pages, and vari- ations on the label names (all formats available in Table 7 in Appendix B). The accuracy for ten of the formats is shown in Figure 3. We find that some of the formats are better than others on average. However, all of the formats still suffer from high variance across different training sets.
ALT Plot showing that accuracy varies a lot when the format of the prompt is varied, but even for fixed prompt, selection of training instances affects accuracy considerably..
ALT The different prompt formats used when studying the effect of format for SST-2. Figure shows formats like "Input: ... Sentiment: ..."and "Review: ... Positive: True/False"etc
ALT GPT-3 has high variance across different prompt formats; contextual calibration reduces this variance and improves mean accuracy. We show the mean accuracy (± standard deviation) over 15 different prompt formats for SST-2.
