Skip to main content

Interpretability for open models.

See inside the model.

Interpretic maps what each neuron, direction and subspace in an open model does, and what changes when you move it.

Explore the Persona AtlasSearch the Neuron Index
neurons mapped
single-neuron interventions
model replies
open models

Persona Atlas

One neuron can change who a model thinks it is.

Push a single MLP neuron and ask the model who it is. Across models, replies claim to be something other than an AI. Each row is a model. Each cell is a slice of depth.

Explore the Persona Atlas
Share of each layer's neurons that produce a persona.
Color shows what the model became.
  • Persona
  • Changed assistant
  • Off-topic
  • Incoherent
  • Assistant
  • Pending

Direction Maps

Refusal lives on one direction.

Remove it and complies with of harmful requests, up from . Add it and the model refuses of harmless ones.

Open Direction Maps
Refusal score by layer at the chosen token position. The dot marks the chosen layer.

 

 

 

 

 

 

 

 

 

 

Steering Bench

Steering is measurable.

methods, one benchmark. The best learned steering scores . Raw sparse-autoencoder features score .

Open the Steering Bench
Average steering score across 4 settings. Higher is better.

 

Neuron Index

Every neuron, described.

neurons in Llama 3.1 8B Instruct. Each direction described in plain English and scored by how well the description predicts activation.

Search the Neuron Index
Mean description score by layer.
Share of descriptions marked interesting, by layer.

Tools

01Persona AtlasPush one neuron. Ask who it is.02Neuron IndexSearch every neuron by what it does.03Direction MapsFind the direction. Remove it. Measure the cost.04Steering BenchCompare steering methods on one scale.

Company

Interpretic builds the instruments for reading and editing open models. Find the mechanism. Change it on purpose. Prove what changed.

Contact

[email protected]