INDEX

Explanations

advice phrases like "you can"

This neuron activates on tokens that mark the start of an answer—e.g. “A:”, “I”, “you”, “think”, “Based”—so it’s detecting the beginning of response passages.

New Auto-Interp

Configuration

Prompts (Dashboard)

24,576 prompts, 128 tokens each

Dataset (Dashboard)

monology/pile-uncopyrighted

Embeds

IFrame

Link

Not in Any Lists

Negative Logits

.},

-1.41

 ogrom

-1.34

För

-1.30

Karakter

-1.27

ſſen

-1.26

 좋은

-1.26

 kulinar

-1.25

diğinde

-1.24

كانت

-1.23

Viitteet

-1.23

POSITIVE LOGITS

 solution

1.42

to

1.34

 This

1.28

1.22

in

1.22

1.18

can

1.14

In

1.09

1.08

Activations Density 0.207%