Closed vasnt closed 3 months ago
This would be a cool feature! I'm not exactly sure how to do it, but I'll try to figure it out!
I used an LLM to tag all the fedidevs accounts! It's not perfect but I think it's pretty good. @vasnt if you see any mistakes let me know, I can manually adjust any account that might be mislabeled!
Demo 🎥
https://github.com/user-attachments/assets/4526ad7e-b878-4d77-89a2-759ef4849700
@anze3db It's really quick & UI is also awesome.
I am sure there will be many incorrect labels. (due to AI hallucinations)
i think you can provide each user provision to provide feedback on correct labelling. & thereafter consolidately, based on users feedback, you/me may review it once in a while & update category/labelling.
I am also interested to know how to use LLM to label.
If it is done by LLM, i would prefer to assign one more task to LLM. You may further divide people into category (or independent filtering) e.g. like:
or you may deploy AI into website, to use AI (in case if doesnt cost more/free) based on one's needs for filtering people..
I'll be happy to contribute in case if i see any incorrect labelling to support the project.You may guide me how to update in case if any incorrect labelling.
I have some very rough UI for fixing the labels, I can give you access to it if you let me know your Mastodon account that you used to log into fedidevs.
I am also interested to know how to use LLM to label.
The source code is here: https://github.com/anze3db/fedidevs/blob/main/accounts/management/commands/classify.py but it's essentially just feeding chatgpt with the account username/bio and asking it to figure it out!
You may further divide people into category (or independent filtering)
For some of these (e.g. CEO) you don't really need the LLM to label it, you can use existing text search for it. example: https://fedidevs.com/?q=CEO . It's not super accurate, but I don't think the LLM would be able to do much better.
unfortunately, i don't have mastodon/fedidevs account yet. If it is must, i ll make one.