Repetition and token leaking
#3
by
Varkoyote
- opened
Hey! I've been testing this model a bit but it seems to quickly fall into repetition for me, and incoherence (repeating an action that happened 4 or 5 messages ago). It also seems to occasionally include "user", "ai" or "assistant" as an end token...
That's quite odd! You're certain you're using ChatML to communicate with the model?