INDEX

Explanations

phrases indicating denial or refusal regarding actions and responsibilities

New Auto-Interp

Top Features by Cosine Similarity

Configuration

ckkissane/attn-saes-gpt2-small-all-layers/gpt2-small_L10_Hcat_z_lr1.20e-03_l11.30e+00_ds24576_bs4096_dc1.00e-05_rsanthropic_rie25000_nr4_v9.pt

Prompts (Dashboard)

36,864 prompts, 128 tokens each

Dataset (Dashboard)

Skylion007/openwebtext

Features

24,576

Data Type

float32

Hook Name

blocks.10.attn.hook_z

Hook Layer

Architecture

standard

Context Size

128

Dataset

Skylion007/openwebtext

Activation Function

relu

Embeds

IFrame

Link

Head Attr Weights

0:0.17

1:0.06

2:0.07

3:0.06

4:0.10

5:0.08

6:0.06

7:0.06

8:0.03

9:0.13

10:0.05

11:0.08

Negative Logits

 Mald

-3.17

 Magnet

-2.96

amphetamine

-2.82

Lib

-2.81

 Miles

-2.79

 Liberty

-2.73

Bridge

-2.69

 Malk

-2.65

 Electro

-2.60

 Silk

-2.56

POSITIVE LOGITS

›

3.01

\":

2.85

utterstock

2.85

 apologise

2.84

kW

2.83

ا�

2.83

�

2.79

 veter

2.74

 Blizz

2.72

 ����

2.69

Activations Density 0.020%

phrases indicating denial or refusal regarding actions and responsibilities

No Known Activations