calculating loss per token

[xiamengzhou]

Hi,

I am looking into the baseline code for calculating the likelihood of the generation, and not sure why in this line it was the loss within this range [-SUFFIX_LEN-1:-1] is considered. I thought it should be loss in [-SUFFIX_LEN:] is considered as it was the first prefix token's loss excluded, so the losses at the last 50 positions should be the losses for the 50 generated tokens.

Maybe I misunderstood tough! Thanks!

[carlini]

I think you might be right. But -- we're not claiming our example here is is the "perfect" way to attack by any means. If you can do better by looking at the last 50 exactly then that would be good to know, and should improve attack success rates.