INDEX

Explanations

stumble upon, stone tool

This neuron simply outputs a unique nonzero code for every input token—i.e. it’s acting as an identity lookup of the token’s vocabulary (BPE) index rather than detecting any linguistic feature.

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

ױ

-1.06

参戦

-1.05

<<<<<<<

-1.00

当時は

-0.98

もありました

-0.97

if

-0.97

we

-0.96

 invader

-0.95

ۋ

-0.95

POSITIVE LOGITS

μως

1.20

ここで

1.14

――

1.13

nextProps

1.13

８

1.12

 découvrir

1.09

 voulu

1.08

αυ

1.06

 تلویزی

1.05

〇

1.02

Activations Density 0.080%