3 ms·
There was a development recently where persuading it to print the string <|endoftext|> would cause it to leak snippets of other conversations. I wonder if this
by vermilingua 3y ago
There was a development recently where persuading it to print the string <|endoftext|> would cause it to leak snippets of other conversations. I wonder if this is a result of that occurring in other conversations that were happening simultaneously, causing “bleedover” of broken behaviour into this one.
- bestcoder69 3y agoThat’s not what’s happening. endoftext is OpenAI’s document splitting token, so when GPT encounters it, it treats it like the end of a chat. So the chat “ends” as far as the model is concerned but the app is still sampling tokens, so GPT just… does its best. The output is similar to when you hit “Generate” with no prompt in OIA playground’s text completion mode. There, you can imagine an invisible endoftext as your prompt.