INDEX

Explanations

engineering concepts

The neuron is highly active on mid‐word subword pieces (wordpiece fragments) rather than whole common words, effectively detecting rare morpheme-like substrings across different tokens.

New Auto-Interp

Configuration

Prompts (Dashboard)

24,576 prompts, 128 tokens each

Dataset (Dashboard)

monology/pile-uncopyrighted

Embeds

IFrame

Link

Not in Any Lists

Negative Logits

 from

-1.66

its

-1.57

-1.37

 erklä

-1.37

on

-1.36

遽

-1.33

 Почему

-1.26

-1.24

 with

-1.23

It

-1.22

POSITIVE LOGITS



1.70

 следующие

1.47

Belgique

1.45

 alemanes

1.44

 ответил

1.40

自然是

1.38

譏

1.35

肯定是

1.34

]_{\

1.34

isième

1.33

Activations Density 0.123%