INDEX

Explanations

punctuation

np_max-act · gemini-2.0-flash

This neuron detects conversation control tokens — message boundary and speaker-role markers (like <|im_start|>, <|im_end|>, "user", "assistant") that indicate turn boundaries in the dialogue.

oai_token-act-pair · gpt-5-mini Triggered by @vetterc0

New Auto-Interp

Configuration

andyrdt/saes-qwen2.5-7b-instruct/resid_post_layer_11/trainer_1

Dataset (Dashboard)

Various

Features

131,072

Data Type

float32

Hook Name

blocks.11.hook_resid_post

Architecture

standard

Context Size

1,024

Dataset

monology/pile-uncopyrighted

Embeds

IFrame

Link

Not in Any Lists

No Comments

Negative Logits

Hou

-0.07

/product

-0.07

Accessory

-0.07

Pa

-0.07

.properties

-0.07

 riff

-0.07

嫩

-0.06

Half

-0.06

_qu

-0.06

Na

-0.06

POSITIVE LOGITS

服役

0.07

meyeceği

0.07

لوح

0.07

outers

0.07

✝

0.07

riday

0.07

郇

0.07

上来

0.07

registry

0.07

imachinery

0.07

Activations Density 0.129%

punctuation

This neuron detects conversation control tokens — message boundary and speaker-role markers (like <|im_start|>, <|im_end|>, "user", "assistant") that indicate turn boundaries in the dialogue.

No Comments

No Known Activations