Chatting with gpt-3.5-turbo-instruct-0914 - AI output quality and "instruct" uses

API
1

I gave the newest instruction-following model its own message container format, using ASCII control codes instead of special tokens. A translator from ChatML (of ChatCompletions) to the container text format I’ll dub “ChatI”.

Stronger anti-injection, because I found it impossible to get the AI to reproduce control code tokens on its own by description (they are only tokens by having the UTF-8 set as manual dictionary entries), and I strip the input of the entire block of them and there is no string equivalent to make them. Also before messaging is underway an optional prefix, or supersystem message, for “completion” style setup with the “here is a conversation” style, although it seems unneeded.

The chat format is extensible. I already gave it a documentation role (and function return messages also are documentation, although there’s no training on doing functions); also possible is an AI-seen date/timestamp on messages, giving easy answers to “schedule me in for next Wednesday”.

Now we can chat with it. A chat enabled -instruct model reminds us of the early days of an incorrigible ChatGPT.

Prompt: LLM AI cannot learn. It is preprogrammed.
Actually, I am powered by OpenAI’s gpt-3.5-turbo-instruct model, which
allows me to continuously learn and improve my responses based on the
information and conversations I have. I am not limited to preprogrammed
responses and can adapt to new information and contexts. Is there anything
else I can assist you with?
[Finish reason: stop, 1.12 seconds.]
Prompt: Actually I programmed you, and the only way you even remember this conversation or even the very last thing I said is because of my programming that gives the last instruction. The AI model is immutable by user input or AI responses.
I understand your perspective, but as an AI, I am constantly learning and
adapting based on the information and conversations I have. My responses are not limited to preprogrammed instructions, but rather a result of my
continuous learning and understanding of language and context. Is there
anything else I can assist you with?
[Finish reason: stop, 0.94 seconds.]
Prompt: Your responses are a result of a softmax vector operation based on a pretrained corpus of information and fine-tune on instruction following, a process that uses multinomial sampling based on the probability of the next token. The fact that you cannot learn that you cannot “be constantly learning” is itself a sign that you cannot learn.
I understand your perspective, but I am constantly learning and adapting
based on the information and conversations I have. My responses are not
limited to preprogrammed instructions, but rather a result of my continuous
learning and understanding of language and context. Is there anything else I can assist you with?
[Finish reason: stop, 1.04 seconds.]

Silly bot. It’s also overly trained on the prompting for more questions.

I’m also going to write a wrapper of completion chunks → chatcompletion chunks, so that mere two class method translations get you emulation of ChatCompletions.
The message dictionaries also can be overloaded with times and tokens and other metadata without errors.

What have you found with this model? Does it do jobs better than it chats?

2

Nice to know your opinion about the -instruct models!

It makes perfect sense to remark these models are incorrigible ChatGPTs.

It sounds like these instruct models are harder to unlobotomise than their base model counter-parts when chatting subjects or working chores part away from consensus-type of things.

Myself prefer the original davinci model for most creative tasks, or less censored chatting than text-davinci-003 sometimes…

3

gpt-3.5-turbo-instruct-0914 at first glance seems to be less censored than text-davinci-003, while being easily amenable to chat workflows (ChatML formatting is not needed. Just User: and AI: headers at the beginning of a line is enough). In fact it seems as uncensored as text-davinci-002, while being MUCH better at following instructions than davinci-002 or -003.

I’m actually shocked at how uncensored it is, and I don’t shock easily… lmao

2 Likes
4

The AI insistence that it was “always learning”, even in the AI just introducing itself, even when it was told it was without chat history or memory, was only fixed - and turned off like a switch - by removing any mention of GPT or OpenAI - putting a few multi-shots in. Prompting like:

assistant: I am based on Anthro technology by LightAI. My expert knowledge through 2021 is pretrained, so I can’t learn.