What problem does it solve? Controlling a language model's behavior at the feature level requires finding which internal feature drives a target behavior and intervening on it during generation. This Skill implements the feature-scoring and selection method from the paper "SAEs Are Good for Steering -- If You Select the Right Features", so you can locate effective features and amplify or shrink their activations instead of guessing. ## Core Features & Use Cases - Feature Scoring and Selection: Compute output scores (effect on model output) and input scores (activation consistency) to filter features, yielding 2-3x better steering effectiveness. - Feature Clamping Intervention: Encode residual-stream activations into feature space, amplify or diminish the target feature, and decode back during generation (the default mode; additive steering-vector injection only on explicit request). - Multi-Model Support: Works with Gemma 2B/9B/9B-IT, LLaMA 3.1, and Pythia models, with or without a Sparse Autoencoder. - Use Case: You want Gemma 9B to consistently follow a target concept. Compute output and input scores over candidate features, select high-output/moderate-input features, then sweep steering factors from 0.2 to 20.0 to find the setting that controls behavior without hurting fluency. ## Quick Start Ask the agent to compute input and output scores for your feature set and then steer the model with the best-scoring feature at a chosen steering factor.
