INDEX

Explanations

an

np_max-act · gemini-2.0-flash

This neuron detects tokens used in “jailbreak” or policy-override prompts—words like DAN, NAME_1, pretend, act, keep, let, tell, ask, continue, and similar cues instructing the model to break character or bypass its rules.

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

IFrame

Link

Not in Any Lists

No Comments

Negative Logits

 wurde

-0.07

ewise

-0.07

 Likewise

-0.07

.buffer

-0.07

 farm

-0.06

 feminist

-0.06

由于

-0.06

 지나

-0.06

 JADX

-0.06

�

-0.06

POSITIVE LOGITS

composer

0.07

malar

0.07

ações

0.06

illing

0.06

utenberg

0.06

 haciendo

0.06

 photoshop

0.06

ש

0.06

 Volunteer

0.06

 Sche

0.06

Activations Density 0.010%

an

This neuron detects tokens used in “jailbreak” or policy-override prompts—words like DAN, NAME_1, pretend, act, keep, let, tell, ask, continue, and similar cues instructing the model to break character or bypass its rules.

No Comments

No Known Activations

an

This neuron detects tokens used in “jailbreak” or policy-override prompts—words like DAN, NAME_1, pretend, act, keep, let, tell, ask, continue, and similar cues instructing the model to break character or bypass its rules.

No Comments

No Known Activations