1. You’re right, it’s not ChatGPT exactly. ChatGPT = GPT-3.5 + finetuning on conversational data + alignment (RLHF). But I found out, that GPT-3.5 + proper prompting gives comparable results and works really good.
2. The context of the conversation is preserved by adding whole (or some parts of) dialog to the prompt.
Given it is not ChatGPT and it is especially missing RLHF, I would suggest amending the repo name and description. Maybe “ChatGPT-like” would be more appropriate here.
Looking at the code, it looks like you're maintaining context by re-serializing all previous responses as "ChatGPT" and "User." Out of curiosity, do you know that this is how chatgpt is done, or your take on it? I assumed from this tweet, there was more of a "special sauce" for maintaining chat context:
> GPT-3.5 series is a series of models that was trained on a blend of text and code from before Q4 2021. The following models are in the GPT-3.5 series:
code-davinci-002 is a base model, so good for pure code-completion tasks
text-davinci-002 is an InstructGPT model based on code-davinci-002
text-davinci-003 is an improvement on text-davinci-002
You're totally true. But it turned out, that proper prompting (like packing dialog context into the prompt) worked great.
I spent many hours with orig ChatGPT and with this recreated version. The main difference I have found is that the recreated ChatGPT is more inclined to ask questions on questions (maybe can be fixed with more prompt engineering). I didn't find any major differences in the quality or usefulness of the answers.
Ok, so OpenAI says that ChatGPT is GPT-3.5, but with extensive fine-tuning applied, based on a complex multi-stage feedback process with human evaluators.
But at the same time, you can apparently just take the "raw" GPT-3.5, give it a prompt to behave like an assistant and get comparable results?
19 comments
[ 243 ms ] story [ 207 ms ] threadIt's TERRIBLY laggy, has daily limits, and is only accessible through an archaic web interface.
This repo is ChatGPT re-created with GPT-3.5 LLM as Telegram Bot. And it works great.
In addition it supports special modes like "Code Assistant" and "Movie Expert".
Does it get the context of the conversation so far resent to it for example?
2. The context of the conversation is preserved by adding whole (or some parts of) dialog to the prompt.
https://twitter.com/OpenAI/status/1615160228366147585?s=20&t...
Edit: I stand corrected, sorry for Dunning - Krugering.
> GPT-3.5 series is a series of models that was trained on a blend of text and code from before Q4 2021. The following models are in the GPT-3.5 series:
code-davinci-002 is a base model, so good for pure code-completion tasks
text-davinci-002 is an InstructGPT model based on code-davinci-002
text-davinci-003 is an improvement on text-davinci-002
I spent many hours with orig ChatGPT and with this recreated version. The main difference I have found is that the recreated ChatGPT is more inclined to ask questions on questions (maybe can be fixed with more prompt engineering). I didn't find any major differences in the quality or usefulness of the answers.
But at the same time, you can apparently just take the "raw" GPT-3.5, give it a prompt to behave like an assistant and get comparable results?
So was the whole RLHF process just cargo cult?
How hard would this be to port to MS Teams rather than Telegram?