INDEX

Explanations

foreign language names

The neuron isn’t looking for any particular word or meaning at all but simply how far into the text it is – its activation rises nearly linearly with the token’s position in the document.

New Auto-Interp

Configuration

Prompts (Dashboard)

24,576 prompts, 128 tokens each

Dataset (Dashboard)

monology/pile-uncopyrighted

Embeds

IFrame

Link

Negative Logits

our

-1.65

 this

-1.61

 here

-1.44

my

-1.44

we

-1.32

 ourselves

-1.24

 yesterday

-1.23

咂

-1.23

 myself

-1.21

 today

-1.18

POSITIVE LOGITS

his

1.61

 после

1.41

 rád

1.35

 setelah

1.30

内心

1.28

因为

1.23

 گیاه

1.22

 njego

1.22

他们

1.21

служба

1.21

Activations Density 0.001%