#interpretability

if anyone sees this who is interested, i'll upload the viewer i built for y'all to play with

posting a pic, because character limit. #machinelearningdevelopers #interpretability

#machinelearning #interpretability

#interpretability #machinelearning #Llama

Llama 3.2 3B, fMRI #MachineLearning #interpretability

llama 3.2 3b, layer 27. lower green = dim barely mattered high pink = it's important AF.

layer 27 vs. layer 28, llama 3.2 3b. look how much of the last layer contributes to the model cognition. striking!

https://t.co/kfOM5G1MqG

Llama 3.2 hidden state, fig 1. layer 28, rotational/grad norm/directional overlay, fig 2/3

llama hidden state neurons for probing. i can't remember what layer but i'm doing them all.

Llama 3.2 3B, RTX 5060Ti. chemical sliders directly modulate the model hidden state during token streaming, just gotta figure out which sliders are wrong