How do you want to poke at the LLM?

nanointerpret

Model: Loading...

Intervention playground

Force the activation of a feature to see how it influences the text generated by the model

How to use the playground

Enter a prompt like I was driving when
Choose the feature that you want to forcefully activate. Open the feature dropdown and scroll or search by ID or title.
Choose how strongly to activate the feature:
- 50% mildly activated
- 100% maximum observed activation
- 200% twice the maximum (may break the model)
Press Generate comparison

Sampling settings

Maximum tokens

Temperature

Top P

Top K

Repetition penalty

Examples

Prompt

Feature

Target activation (% of feature max)

50% 100% 150% 200%

Unmodified

Intervened

Filters

Category

Activation count

Sort

Loading activations... High quality Activations

Page