INDEX

Explanations

instances where the phrase "can't" or "t" followed by a high activation word are used

New Auto-Interp

Top Features by Cosine Similarity

Configuration

ckkissane/attn-saes-gpt2-small-all-layers/gpt2-small_L2_Hcat_z_lr1.20e-03_l11.00e+00_ds24576_bs4096_dc1.00e-06_rsanthropic_rie25000_nr4_v4.pt

Prompts (Dashboard)

36,864 prompts, 128 tokens each

Dataset (Dashboard)

Skylion007/openwebtext

Features

24,576

Data Type

float32

Hook Name

blocks.2.attn.hook_z

Hook Layer

Architecture

standard

Context Size

128

Dataset

Skylion007/openwebtext

Activation Function

relu

Embeds

IFrame

Link

Head Attr Weights

0:0.06

1:0.02

2:0.11

3:0.08

4:0.13

5:0.11

6:0.05

7:0.03

8:0.14

9:0.11

10:0.07

11:0.04

Negative Logits

tailed

-1.28

bush

-1.26

Kub

-1.16

quartered

-1.15

rant

-1.13

 Heads

-1.11

interstitial

-1.11

Vers

-1.11

 pitted

-1.11

warts

-1.11

POSITIVE LOGITS

ل

1.50

ック

1.44

ICLE

1.42

ゼウス

1.32

ы

1.30

TAG

1.30

EEEE

1.30

ivably

1.29

ustain

1.28

adena

1.27

Activations Density 0.027%

instances where the phrase "can't" or "t" followed by a high activation word are used

No Known Activations