removing bnb config

quantization trial

adding another variation of Admitted.

Admitted bad words

adding logging + some bad words

adding back truncation

lowering temp to 0.7

taking out .to(cuda)

back to .generate. forgot about returning gen ids

bug + only new tokens

bugfix

pipeline trial

torch.cuda.empty_cache()

taking out truncation of input_ids

do_sample=True, and model.to('cuda')

truncating earliest tokens if longer than 4092

arg name change, Abort

adding periods back, adding Aborted.

taking out . after Abort etc

adding Ab ort . to bad words

adding additional bad words ids

banning Admitted and Abort

banning token 13

returning generated ids

getting generated_ids[0]

back to normal, removed additional bad words argument

adding intros k v l. to bad words (for testing)

using .generate, returning ids, custom bad_words

adding new bad word

returning ids

printing generated ids

testing .generate instead of pipeline

testing custom logit processor for bad words

temp=2! top_k 40, 1 ret seq

num ret seqs 2

testing high temperature

adding second way of describing bad word (*

increase temp, add top_k=20

XXS comment

temperature 0.7

only return new generated tokens, bad_words = comments

base llemma-7b and custom handler.py