INDEX

Explanations

math and code punctuation

The neuron fires on non-natural “token junk” such as URL fragments, numeric strings, email addresses, or other technical/code-like identifiers.

New Auto-Interp

Configuration

Prompts (Dashboard)

392,802 prompts, 256 tokens each

Dataset (Dashboard)

monology/pile-uncopyrighted

Embeds

IFrame

Link

Not in Any Lists

Negative Logits

。『

0.88

 Един

0.80

 synthesized

0.80

鹖

0.79

 ██

0.77

 کی۔

0.77

❤

0.76

 Официа

0.75

।"

0.74

 নীলন

0.73

POSITIVE LOGITS

iu

0.88

audience

0.86

hi

0.84

isu

0.82

hips

0.82

0.81

util

0.79

aa

0.79

invited

0.79

ie

0.78

Activations Density 0.001%