INDEX

Explanations

results of recent work

This neuron detects title‐ or headline‐style text, i.e. sequences of words or symbols that form document titles, section headers, or other high‐level headings.

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

for

-1.81

if

-1.62

 then

-1.45

 when

-1.43

 again

-1.30

 unless

-1.30

is

-1.19

↵↵↵↵↵↵↵↵↵↵↵

-1.14

to

-1.13

 after

-1.13

POSITIVE LOGITS

 美味

1.20

the

1.20

ął

1.19

 aceptas

1.16

 choisi

1.15

 byproduct

1.12

 быстро

1.11

 parfois

1.10

 piele

1.09

 fotografías

1.08

Activations Density 0.246%