Layer Analysis

Not all layers are equally effective for steering. PSYCTL includes tools to analyze which layers provide the best separation between personality-exhibiting and neutral activations.

CLI Usage

psyctl layer.analyze \
  --model "google/gemma-3-270m-it" \
  --layers "model.layers[*].mlp" \
  --dataset "./dataset/steering" \
  --method svm \
  --top-k 5

Analysis Methods

SVM Analyzer

Trains a linear SVM at each layer to classify positive vs neutral activations:

• Score: Overall separation quality (higher is better)
• Accuracy: Classification accuracy
• Margin: SVM margin (larger = more robust)

Consensus Analyzer

Combines multiple analysis methods for more robust layer selection.

Layer Patterns

Wildcard patterns for targeting groups of layers:

Pattern	Matches
model.layers[*].mlp	All MLP layers
model.layers[5:15].mlp	Layers 5-14
model.layers[::2].mlp	Every other layer

Interactive Notebook

Try the 05_layer_analysis notebook for a visual walkthrough with charts.