Erronous responses

#1
by colourspec - opened

When using this model unquantized in aphrodite with ChatML, saw few garbled responses.

In one of the responses it completely jumbled all the words of a sentence.
In a couple of responses it would randomly insert alpaca format, eg. first half of the response being correct, but second half will be: ### Instruct and a random different story.

Using temp as low as 0.7 didn't help either.
Also it seems to be sloppy or gets sloppier with responses.

The Theta model didn't have these problems.

Sign up or log in to comment