incorrect neuronpedia reference for few-shot sample in intruder_prompt.py #154
Description
delphi/scorers/classifier/prompts/intruder_prompt.py includes some samples for few-shot prompting and most of them have a helpful comment pointing to the original neuronpedia link where that example came from. however, i think DSCORER_EXAMPLE_THREE has the wrong link attached, as i was not able to find it in the attached link and it matches the link for the second example. probably a copy-paste error?
# https://www.neuronpedia.org/gpt2-small/8-res-jb/12654
DSCORER_EXAMPLE_TWO = """Examples:
Example 0: enact an individual health insurance mandate?âĢĿ, Pelosi's response was to dismiss both
Example 1: climate, TomblinâĢĻs Chief of Staff Charlie Lorensen said.Ċ
Example 2: no wonderworking relics, no true Body and Blood of Christ, no true Baptism
Example 3:ĊĊIt has been devised by Director of Public Prosecutions (DPP)
Example 4: and fair investigation not even include the Director of Athletics? · Finally, we believe the
"""
DSCORER_RESPONSE_TWO_COT = "I can see that there are several examples that have the word 'of' before a capital letter. The intruder is 0 because it does not."
DSCORER_RESPONSE_TWO = "[RESPONSE]: 0"
# this is the same link as up above, i wasn't able to find these words on the page but i was able to find the examples above on it
# https://www.neuronpedia.org/gpt2-small/8-res-jb/12654
DSCORER_EXAMPLE_THREE = """Examples:
Example 0: Climbing
Example 1: running
Example 2: swim
Example 3: eating
Example 4: cycling
"""
DSCORER_RESPONSE_THREE_COT = "All examples are related to activities, the first 3 and the last one being about sports and physical activities. Eating is not a sport or physical activity so it is the intruder."
DSCORER_RESPONSE_THREE = "[RESPONSE]: 3"
i would appreciate if it could be updated to the correct link, but if that is no longer available then it's probably better to just remove it altogether