SAE & Interpretability

Orrery treats Sparse Autoencoders as ordinary vector collections. Each SAE feature is a row of the decoder matrix — a vector in the model’s space — so a feature set can be downloaded, projected, clustered, and labelled by the same pipeline as any dataset, then explored as a galaxy where entering a prompt highlights the features it activates.

A Gemma Scope 2 SAE feature space, projected and labelled by the collection pipeline

Two model families are supported for live inference: Gemma 3 (Gemma Scope 2 SAEs, with labels and activation examples from Neuronpedia) and Qwen (Qwen-Scope, label-free), with from-scratch JumpReLU/TopK implementations in plain PyTorch. Any feature can be inspected in a dashboard — top/bottom logits, activation histograms, similar features, activating examples — and injected into the model with additive steering (ActAdd) in a chat interface.

Steered chat: with the "poetry" feature injected, Gemma-3-4b-it answers "How do you feel?" in a poetic register

In this section

SAE Storage & API — DuckDB schema, ingestion, GraphQL API
Data Pipeline — SAE download-to-ingestion, Neuronpedia labels
Live Inference API — prompt activations, steering, streaming chat

See also the case studies, where the steering backend is validated against a published refusal-direction result.