INDEX

Explanations

social media handles or usernames containing specific keywords or symbols

oai_token-act-pair · gpt-3.5-turbo

specific repeated symbols or characters

oai_token-act-pair · gpt-4o-mini Triggered by @bot

New Auto-Interp

Top Features by Cosine Similarity

Comparing With GPT2-SMALL @ 5-res-jb

Configuration

jbloom/GPT2-Small-SAEs-Reformatted/blocks.5.hook_resid_pre

Prompts (Dashboard)

24,576 prompts, 128 tokens each

Dataset (Dashboard)

Skylion007/openwebtext

Features

24,576

Data Type

torch.float32

Hook Point

blocks.5.hook_resid_pre

Architecture

standard

Context Size

128

Dataset

Skylion007/openwebtext

Hook Point Layer

Activation Function

relu

Embeds

PlotsExplanationShow Test FieldDefault Test Text

IFrame

Link

Not in Any Lists

No Comments

Negative Logits

userc

-0.76

 cryst

-0.75

pex

-0.74

 livest

-0.74

 destro

-0.69

nut

-0.66

 advoc

-0.66

aturdays

-0.66

 compr

-0.65

 territ

-0.65

POSITIVE LOGITS

————

0.80

 Protesters

0.70

 Hide

0.67

 Hear

0.67

 Recap

0.65

 Written

0.65

 Chapter

0.65

 Rohingya

0.64

————————

0.62

Rat

0.62

Activations Density 0.078%

Top

Activations

—

15.22

0.↵
<|endoftext|>vQHw — NC State Baseball (@NCStateBaseball

—

14.63

Lives.<|endoftext|>eso1338 — Science Release↵
↵
Bizarre Alignment

—

14.31

: We don<|endoftext|>urden sharingâĢĿ — and instead encourage an approach that builds on

—

14.01

Ļs worst<|endoftext|>2DejiV — FootballScoop Staff (@FootballS

—

13.61

<|endoftext|>EgKQbnxU — Toe in the Water (@toeint

—

13.60

/DvrzWiOeYD — Kurt Schlichter (@KurtSch

—

13.44

luck has run<|endoftext|> important benefitâĢĿ — employment — there must be a better reason

—

12.91

<|endoftext|>u8yDEV3HeP — Sky Sports Statto (@SkySportsSt

–

12.82

use of plastic<|endoftext|>Ļs Obscene – Politics, Culture, and Comedy in America

—

12.77

t find<|endoftext|> (WJZ) — A nine-month-old girl was

—

12.64

the 2015 NFL<|endoftext|>.#GTAOnline — Fun 2 (@TezFunz2

—

12.64

/QGMGSb8Wjl — Tolarian Community (@TolarianCollege

—

12.49

home.↵
<|endoftext|> taxes.âĢĿ — Barack Obama, September 2008↵
↵
Mr

—

12.46

.↵
↵
<|endoftext|>sem Youssef — often described as EgyptâĢĻs version

—

12.45

/dke2syOrhj — dan holler (@danholler)

—

12.35

xAfzqFgnym — Noah Jefferson ï£¿ (@noah

—

12.23

treat patients<|endoftext|>âĢĻt be one — you talked about âĢľthemâĢĿ

—

12.17

ization and job<|endoftext|> $25 an hour — between $31,200 and $52

—

12.05

Z95nQKCx3D — NASA Watch (@NASAWatch) June 13

—

12.00

iWoPjG6yBs — James Woods (@RealJamesWoods)

—

11.61

you<|endoftext|>ids in comparison.âĢĿ — Heinrich Heine, Lutet

—

11.36

ums in Brazil<|endoftext|> a person in need — âĢľthe least of my brothers and

—

11.17

<|endoftext|>↵
↵
8 p.m. — The Big Bang Theory/Kevin Can Wait

–

10.78

ya-san.↵
↵
Source Article – Anime.dkmt↵
↵
Source

—

10.70

'll ever<|endoftext|>-mansions be damned — toward more compact living: The micro-

—

10.40

Orange<|endoftext|>year-old girlâĢĿ—on the complaint against Donald, which was

—

9.51

by statistics portal<|endoftext|> giving increased in 2013 — the first time since the 2008 recession.

—

9.23

0↵
↵
10 p.m. — Blue Bloods↵
↵
Saturday↵
↵

–

8.32

article is a<|endoftext|> that open the risk – they want to be online on Facebook,

—

8.07

MMXwvthYlS — Ally Maynard (@missmayn)

—

7.69

call it âĢľtechnoâĢĿ — has graduated from the kiddie table

7.65

Voice-Over Role (Television) - Young Actor" category in the 34th

–

6.60

<|endoftext|>Federal<|endoftext|> transgender, gay, lesbian – people who look different. IâĢĻ

–

6.56

material,<|endoftext|>e., the named defendants – are the ones who utilized the internet access

–

5.90

Jeff Garlett in round 19, 2010 – suggests he's on the right track.

–

5.79

's streaming in May New movies for May – "Hook" (1991) --

)—

5.70

iliation Act, or JGTRRA)—were allowed to expire. In addition,

—

5.60

season. Records go back to 1872 — so that would be a big deal.

–

4.89

<|endoftext|>. Who knows what he will say – whether he will renew his attack on secular

–

4.07

novices.↵
↵
Photo Editor – Photo Editor is a simple and easy application

—

4.07

not see the most high-value prisoners — including Abu Zubaydah and Khalid

3.76

<|endoftext|>acist. A drama and cinema buff - early pictures show him as a rak

—

3.60

that researchers also collected audio from news broadcasters — but again, this meant selecting only the

—

3.38

, thatâĢĻs not the case—I came across a Conference Diversity Distribution Calculator

2.83

reflected the attitude of their manager at City - perpetual motion, positive and bristling with

—

2.23

its units are structured for years to come—maybe even as much as the Yom

1.98

leader in progress since then. Five states -- from Massachusetts to Iowa -- have already beat