3 ms·
> the middle layers of the Transformer that are affected most by RL post-training This is where you should expect most change in models. The beginning layers n
by godelski 4mo ago
> the middle layers of the Transformer that are affected most by RL post-training
This is where you should expect most change in models. The beginning layers need to embed while the later layers will reform the result to the final conditions. The middle is what does all the untangling.