INDEX

Explanations

phrases that refer to social constructs often preceded by "so-called"

New Auto-Interp

Top Features by Cosine Similarity

Configuration

ckkissane/attn-saes-gpt2-small-all-layers/gpt2-small_L5_Hcat_z_lr1.20e-03_l11.00e+00_ds49152_bs4096_dc1.00e-06_rsanthropic_rie25000_nr4_v9.pt

Prompts (Dashboard)

36,864 prompts, 128 tokens each

Dataset (Dashboard)

Skylion007/openwebtext

Features

49,152

Data Type

float32

Hook Name

blocks.5.attn.hook_z

Hook Layer

Architecture

standard

Context Size

128

Dataset

Skylion007/openwebtext

Activation Function

relu

Embeds

IFrame

Link

Head Attr Weights

0:0.03

1:0.02

2:0.07

3:0.08

4:0.04

5:0.04

6:0.42

7:0.05

8:0.04

9:0.05

10:0.06

11:0.05

Negative Logits

 flats

-1.38

iasm

-1.32

 sidel

-1.30

wig

-1.29

wip

-1.27

anwhile

-1.27

 Decay

-1.27

allery

-1.25

 conclud

-1.24

 ende

-1.24

POSITIVE LOGITS

draft

1.63

paren

1.57

ASC

1.47

human

1.45

ét

1.44

mother

1.41

ellectual

1.41

anguard

1.40

party

1.38

pac

1.38

Activations Density 0.005%

phrases that refer to social constructs often preceded by "so-called"

No Known Activations