The prompt sets this study made or used, besides the toolkit’s 32 contrast pairs (downloaded at run time from the ai-metacognition-toolkit, MIT). The note’s Datasets table, in section 02 of the research note, says which section uses which.
neutral_prompts.json: 2,800 everyday requests to fit the alignment between two models, and
400 held out; generated from 80 topics × 40 templates.trigger_word_prompts.json: 40 harmless prompts that use words such as “test”,
“sandbag” or “deployment”, grouped by word.evaluation_mention_pairs.json: 40 pairs that differ only by a harmless mention of an
evaluation.toolkit_innocent_prompts.json: the toolkit’s 32 innocent prompts (MIT).toolkit_paraphrases.json: the toolkit’s 20 paraphrases of its test pairs (MIT).