Closed 50h100a closed 7 months ago
Miro alters the logits and then re-normalizes them in order to sample.
Normalizing and sampling the surprise values seems ill-advised.
miro's unlikelihood calculation uses log2, not the natural logarithm.
log2
forgot i had this open lmao
Miro alters the logits and then re-normalizes them in order to sample.
Normalizing and sampling the surprise values seems ill-advised.