INDEX
Explanations
punctuation
The neuron fires on the little “Q:” / “A:” labels that mark question‐and‐answer blocks, i.e. it picks up the section markers (especially the “A” of an answer).
New Auto-Interp
Negative Logits
مث
-0.07
متح
-0.06
cki
-0.06
сторон
-0.06
vista
-0.06
保
-0.06
sweeping
-0.06
рос
-0.06
�
-0.06
گیرد
-0.06
POSITIVE LOGITS
átní
0.07
están
0.07
Opaque
0.06
ichage
0.06
recognition
0.06
Appe
0.06
icia
0.06
slump
0.06
γ
0.06
commend
0.06
Activations Density 0.028%