A tiny request for v2

#6
by mammour - opened

Hey !

First, I wanna thank you for all your work, prior and current, Helixnet is awesome. I've been testing some stuff and the most practical and useful idea I got from my twistings and twirlings was kinda evident, do some cycles (actor -> [critic -> regen] n times).
Results are astounding 😍

I wonder if that is due to the training data or something else, but all of the regenerator's answers start with "REGENERATED ANSWER: ". It was going crazy after multiple cycles.

Luckily the problem is kinda simple, so the solution too, a small change the regen prompt was enough to dodge any bothersome programing :
prompt_regenerator = f"SYSTEM: {system_prompt}\nUSER: {user_input}\nRESPONSE: {response_actor}\nCRITIQUE: {response_critic}\nREGENERATOR: REGENERATED ANSWER: ", it gives clean answers and no side effects, but it feels a bit 'ugly'.

I did read that you would work on some v2. Could this be possible for the regen model to not output this ? It's solely for the sake of soothing wicked minds like mine, if that change the way everything can work properly or if it's too much work don't mind it πŸ€—

Take care 🫢🏽

Hey @mammour ,
Thank you for your message. Yes I know! It's not ideal -- I wrote about this on my Twitter and Reddit. Thank you for finding a fix!
This is due to dataset being contaminated with REGENERATED ANSWER. As you can understand, I'm doing this just by myself purely out of passion. My day job is pretty intense, so I only allocate a few hours of my time every week to training open source LLMs. I need to clean up the dataset -- but I also have started creating the V2 dataset and seems like with better prompting, I have managed to get rid of that "REGENERATED ANSWER" part. I will fix this issue in V2.
I'm pumped to hear that you're getting good results with HelixNet! Keep those feedback coming -- I like to hear from the users.
Migel

That is exactly what I had in mind, I'll keep pushing any relevant feedback !
Thanks for crediting me πŸ€—

migtissera changed discussion status to closed

Sign up or log in to comment