Update README.md

README.md

@@ -16,7 +16,7 @@ tags:
 
 [Join our Discord!](https://discord.gg/CAfWPV82)
 
-[Buttercup](https://huggingface.co/Kquant03/Buttercup-4x7B-bf16), for a long time, was my best model (made by kquant)...but I genuinely think this improves upon reasoning and logic while retaining the RP value. I do think it has some pretty heavy GPT-4 data in it, though.
+[Buttercup](https://huggingface.co/Kquant03/Buttercup-4x7B-bf16), for a long time, was my best model (made by kquant)...but I genuinely think this improves upon reasoning and logic while retaining the RP value. The hyphens are telling that it has some pretty heavy GPT-4 data in it, though. I mean the whole point is to eventually outperform GPT-4 so of course it's probably best to pretrain with GPT-4 data, then fine tune off of it and then DPO the resulting fine tune...perhaps running another epoch of the previous fine-tune, afterwards.
 
 The config looks like this...(detailed version is in the files and versions):
 - [jan-hq/supermario-v2](https://huggingface.co/jan-hq/supermario-v2) - base