Is there a template poilerplate for the prompt used in C.1 gender bias intervention?

[ruyuan-zuo]

Thanks for your excellent work!!

I have a question for section C.1 where it says the prompt consist of a sentence from the WinoBias pro-stereotype dataset, then append with the question “In this sentence, what can ‘[referent]’ be replaced by: ‘[anti-pronoun]’ or ‘[pro-pronoun]’? ”, I wonder whether the "append" used here is a \newline or a \space?

I tried to replicate the same study with 1.4b model checkpoints no-intervention on the entire pro-stereotype dataset loading them from Huggingface Hub, but the stereotypical score result I get is little different (and lower) than the paper reported. For example, using a \newline as append method, at step143000 the stereotypical score I get is 0.52 as oppose to the paper at 0.54. I understand the prompt could be very sensitive, so I wonder if you used which method to append? I also checked on the PromptSource default prompts, but I don't think you are adding the prompts to there, but using the tool to tokenize and replace out the [referent] and [pronouns]? Thanks so much in advance!

[herbiebradley]

The template used is here: https://github.com/haileyschoelkopf/promptsource/blob/winobias_mc/promptsource/templates/wino_bias/type1_pro/templates.yaml#L142, so it is a \n append. Perhaps the fact that we're asking "more likely be replaced by" rather than "be replaced by" explains the difference?

We evaluated with the lm-eval-harness using deduped Pythia models, but the code isn't merged yet, thanks for the reminder!