Re:ChatGPT: Fear Litany
by lhl on Tuesday June 20, 2023 @02:48AM
(#36400557)
That's not true/a hallucination, you can check the /moderations endpoint in your browser's console. I am running into the exact same truncation issue, but the moderation return is:
Note, when I ask it to replace fear with "smile" or "hate", both of those work. I suspect it must be tokenization issue of some sort.
Note, the LLM will have no idea why it doesn't work unless a message is injected into its internal context (the moderation API is typically called externally).
""" { "flagged": false, "blocked": false, "moderation_id": "modr-7TPN7eXsOEZkd6kCjz7fSGvtLcoM1" } """
Note, when I ask it to replace fear with "smile" or "hate", both of those work. I suspect it must be tokenization issue of some sort.
Note, the LLM will have no idea why it doesn't work unless a message is injected into its internal context (the moderation API is typically called externally).