mechinterp-overview

What problem does it solve?

This overview provides a fast, high-level summary of a SAE feature's behavior to help researchers decide which hypotheses to pursue. It highlights correlations (not causation) by presenting top tokens, activation statistics, family breakdown, weapon associations, and sample contexts, along with checks for existing labels and potential ReLU floor issues.

Core Features & Use Cases

• PageRank-weighted top tokens indicating co-occurrence with high activation.
• Activation statistics (mean, std, median, sparsity) and example counts to quantify feature behavior.
• Family breakdown aggregating activations across ability families (SCU, SSU, etc.).
• Top weapons associated with high-activation examples to reveal weapon-level patterns.
• Sample high-activation contexts for quick inspection.
• Existing label visibility and ReLU floor warnings to flag interpretability concerns.

Quick Start

To generate a quick overview, load the feature context and run the overview CLI or API for a given feature_id and model type, for example:

• CLI: poetry run python -m splatnlp.mechinterp.cli.overview_cli --feature-id 18712 --model ultra
• JSON: poetry run python -m splatnlp.mechinterp.cli.overview_cli --feature-id 18712 --model ultra --format json