Post

aria
@aurelium.me

if you want you can just try using a model post-trained mostly by just SFT'ing on hundreds of millions of frontier model replies. it's like 80% of the models on huggingface generally speaking, they suck and barely work. you need extensive RL to make modern models work, there's no way around it

SE Gyges@segyges.bsky.social

in case anyone doesn't know this the thing about open weights models being made by stealing outputs from big players is almost entirely made up to justify regulations locking openai, anthropic and for some reason google into a cartel position

0 reposts6 likes

aria
@aurelium.me

SFT'ing on the outputs of closed models can be useful for getting past the difficult cold-start phase, and serve as an "eh-it's-fine" level replacement for a more disciplined synthetic data flywheel but it cannot really bring you meaningfully up to or past the frontier, afaict