INDEX

Explanations

attends to "how," "what," and "who" from related quantifying tokens

oai_attention-head · gpt-4o-mini Triggered by @bot

New Auto-Interp

Top Features by Cosine Similarity

Comparing With GEMMA-2-2B @ 10-gemmascope-att-16k

Configuration

google/gemma-scope-2b-pt-att/layer_10/width_16k/average_l0_70

Prompts (Dashboard)

36,864 prompts, 128 tokens each

Dataset (Dashboard)

monology/pile-uncopyrighted

Features

16,384

Data Type

float32

Hook Name

blocks.10.attn.hook_z

Hook Layer

Architecture

jumprelu

Context Size

1,024

Dataset

monology/pile-uncopyrighted

Activation Function

relu

Embeds

IFrame

Link

Not in Any Lists

No Comments

Head Attr Weights

0:0.08

1:0.08

2:0.13

3:0.05

4:0.03

5:0.01

6:0.11

7:0.47

Negative Logits

 ModelExpression

-0.37

 aDecoder

-0.35

IBOutlet

-0.33

Geplaatst

-0.33

ngdoc

-0.32

DebuggerNonUser

-0.29

@"/

-0.29

ionario

-0.29

Parcelize

-0.29

 Wikimedijinoj

-0.27

POSITIVE LOGITS

脚注の使い方

0.37

moiselle

0.30

iocene

0.30

kozó

0.30

 drifts

0.29

norsk

0.29

ddots

0.29

 zitate

0.29

 Bucure

0.29

 Squeeze

0.29

Activations Density 0.554%

attends to "how," "what," and "who" from related quantifying tokens

No Comments

No Known Activations