@ChangJonathanC yep it’s a kind of MLP layering technique that can be traced back to 2019, there’s also another model called Yi that used the same technique
I also found another interesting paper called ‘transformer layers as paints’
@_lyraaaa_ it's already there in 9b-base (14-16/20), so it predates post-training and survived the copy plus all of 27b's training
4b ids don't carry over: 4b and 9b aren't index-aligned (0/32 layers)
nothing about the growth in qwen's model cards or blogs that i could find
@_lyraaaa_ We picked it out of ~1M neurons: gradient screen (which neurons raise log p of hostile vs neutral replies), then push only the top 30 plus same-layer random controls. blind judges (llama 3.1 8b + claude haiku). random neurons: max 1/20