INDEX

Explanations

refusing harmful or unethical projects

The neuron fires strongly on tokens used in subjective first‐person statements or personal commentary (e.g. “This is the kind of story I would have wanted…,” “thank you,” “excited,” etc.), so it’s detecting personal opinion or endorsement language.

New Auto-Interp

Configuration

Prompts (Dashboard)

392,802 prompts, 256 tokens each

Dataset (Dashboard)

monology/pile-uncopyrighted

Embeds

IFrame

Link

Not in Any Lists

Negative Logits

 swollen

0.56

Wei

0.55

 Liang

0.53

过

0.53

 Rayleigh

0.52

 faiblement

0.52

 Subsequent

0.51

qua

0.51

POSITIVE LOGITS

 proyecto

2.28

 проекта

2.20

 projektu

2.17

 projekt

2.16

プロジェクト

2.16

 프로젝트

2.16

 project

2.14

 progetto

2.03

 projects

2.02

 projek

2.00

Activations Density 0.868%