"Too many requests in 1 hour. Try again later."
The model is very good, and i got very good results running it on Colab premium (and even local, 4bit quantized). But it is spitting out ChatGPT/OpenAI messages. The weirdest one was "too many requests in 1 hour. Try again later.", as well as some OpenAI messages about policies. The message changes after i changed my prompt to something completely different, so it has definitely been trained with those messages.
I don't have the prompt that resulted on this message, but i can confirm it happened multiple times while running it on Colab premium with 'text-generation-webui'.
I suppose that is one of the drawbacks on training such models using ChatGPT :D
Yeah I'm afraid so! That's hilarious about the "too many requests message", I've not seen that one before :)
Whenever an organisation or group releases a model like this, other people then try to create "unfiltered" versions of it. What they mean by "unfiltered" is actually itself filtering - they edit the source dataset and try and remove all the messages with stuff like "I'm sorry, but as a Large Language Model I can't tell you how to murder puppies" or whatever. They aim to leave all the usable content, but with none of the sentences where it said it was an LLM or why it couldn't answer in full.
There were several such releases for Vicuna, and I expect before long someone will do the same for StableVicuna. I know that at least one person is already working on that for WizardLM for example (@ehartford )