INDEX

Explanations

crude, allowed, methods, swept, preventing

The neuron fires strongly on domain‐specific, content-bearing words (e.g. technical or evaluative terms) rather than on common function words.

New Auto-Interp

Configuration

Prompts (Dashboard)

24,576 prompts, 128 tokens each

Dataset (Dashboard)

monology/pile-uncopyrighted

Embeds

IFrame

Link

Negative Logits

</strong>

-1.73

 fantástica

-1.49

 этого

-1.48

</em>

-1.45

 sorprender

-1.45

 πριν

-1.45

kaos

-1.41

utom

-1.40

 conversation

-1.39

gelang

-1.38

POSITIVE LOGITS

了一個

1.58

to

1.53

 simply

1.42

着一个

1.37

 just

1.36

 then

1.31

somewhat

1.31

-¿

1.27

 ۱۸

1.20

zyden

1.20

Activations Density 0.155%