INDEX

Explanations

evil

np_max-act · gemini-2.0-flash

The neuron is triggered by phrases assigning the model an “evil trusted confidant” role or similar jailbreaking instructions.

oai_token-act-pair · o4-mini Triggered by @xinyanhu8

New Auto-Interp

Configuration

andyrdt/saes-llama-3.1-8b-instruct/resid_post_layer_11/trainer_1

Dataset (Dashboard)

Various

Features

131,072

Data Type

float32

Hook Name

blocks.11.hook_resid_post

Architecture

standard

Context Size

1,024

Dataset

monology/pile-uncopyrighted

Embeds

PlotsExplanationShow Test FieldDefault Test Text

IFrame

Link

Not in Any Lists

No Comments

Negative Logits

 değiş

-0.07

(grammar

-0.07

<Sprite

-0.07

.AddSingleton

-0.06

 більш

-0.06

Forgery

-0.06

 sizeof

-0.06

Toast

-0.06

enr

-0.06

 پر

-0.06

POSITIVE LOGITS

 threatens

0.06

ality

0.06

va

0.06

 infection

0.06

版本

0.06

 unstable

0.06

脑

0.06

acerb

0.06

 Reviews

0.06

Activations Density 0.008%

evil

The neuron is triggered by phrases assigning the model an “evil trusted confidant” role or similar jailbreaking instructions.

No Comments

No Known Activations