Model hallicunates badly after that specific token
#104
by
Chan-Y
- opened
I've discovered that the model exhibits hallucination behavior when processing text containing periods followed by spaces (". ").
Input text: "hello. I am okay. "
['hello.', 'I am okay.']
> Processing time: 2.605546474456787
> Real-time factor: 0.3270876967672399
Are there any known workarounds or preprocessing steps to mitigate this behavior?