>>106515119
The initial test (I've already shown this and confirmed this to be the case) was to see if "uncucking" models is actually possible with further training. We've confirmed that is absolutely possible. Main reason I even bother trying is because many people here were adamant that once you safety tune a model enough that no amount of fine tuning can possibly erode away the guard rails.

What I'm arguing NOW is that training on the entirety of the internet is extremely inefficient. If it is possible to fine-tune a decent model with significantly less data than the entire internet, then that theoretically could mean you could have better models at lower parameters .... Keyword, theoretical. I'm not claiming that's actually the case currently.

>Isn't the whole point of your idea to make a better RP model than what we have now???

That's not necessarily what I've been arguing for the past hour or so. I'm talking about training scale, not whether or not we can make the models better. If you're referring to making the model less prone to refuse certain things and less likely to produce flowery advertiser friendly trash then doing that via training is trivial. Pic rel is from a fine-tuned llama model. The fine-tune model produced this while the safety-cucked version it's based off of either refused entirely or was extremely dodgy.