Closed saliksik closed 8 months ago
+1 This would be very much appreciated if you open-sourced the backend. Is there any reason why it wasn't released together with the frontend originally @shobrook ?
Additionally, these constants do nothing as it stands. Everything happens on the backend:
const EDIT_PROMPT_PARAMS = {
model: "code-davinci-edit-001"
};
const COMPLETION_PROMPT_PARAMS = {
model: "text-davinci-003",
max_tokens: 500,
temperature: 0.2,
top_p: 1,
presence_penalty: 0,
frequency_penalty: 0,
best_of: 1,
n: 1,
stream: false
};
I haven't read the source code in its entirety, but just looking at your comments, "code-davinci-edit-001" and "text-davinci-003" appear to be OpenAI models that are known to only be available through OpenAI's API.
It might be possible to refactor the code to use other models that are downloadable, but the downloadable models may not be as good in terms of high level comprehension. You can get help from Adrenaline, ChatGPT, or Copilot with setting up the downloadable models such as GPT 2. It is over 10 GB though. Once you get a code model and an natural language model setup, you can go to the facehugger community webpage and explore options to possible switch out with other models.
However, I suspect that the downloadable models won't be nearly as good for coding because they haven't had acceess to as large of a training set for coding, and/or they are a previous generation of AI model.
With the release of cool open source models such as Code Llama, I was wondering if this changed
Right now the API is closed-source but we will explore exposing some endpoints in the future (specifically for generating visualizations).
And yeah nothing beats GPT-4 right now.
Hi Sir!
Do you have plans open sourcing the API endpoint of Adrenaline? https://rubrick-api-production.up.railway.app/api/
Thank you.