Tainted Training Data?
#13
by
mrjackspade
- opened
I've been using the model to generate a text chat. Just standard format.
Bob:
Alice:
The last 4 chats I generated like this, the "Alice" character would start saying they were an AI, and asking the "Bob" character if they needed assistance with things. Basically all the things you would expect from an instruct model, which has a pretty serious negative impact on its value as a text completion model.
I'm getting this behavior with no references to AI at all in the context, and its pretty consistant.
Has anyone else seen this behavior?
Edit: I should also add that the model is generating <|eot_id|> characters as well. Not consistently, just seemingly randomly at the end of messages.