Skip to content

SPEC.md

Latest commit

 

History

History
31 lines (18 loc) · 1010 Bytes

SPEC.md

File metadata and controls

31 lines (18 loc) · 1010 Bytes

Interpretability Specification

Version: v1.2.3 | Status: Active | Last Updated: March 2026

Overview

Provides Sparse Autoencoder (SAE) tools for mechanistic interpretability of neural networks. Trains SAEs on activation data to discover interpretable sparse features and analyze activation patterns.

Functional Requirements

  1. Sparse Autoencoder training with L1 sparsity penalty on activation data
  2. Feature analysis: top activated features, sparsity metrics, and activation frequencies
  3. Configurable d_input, d_features, learning rate, and L1 penalty for SAE architecture

Interface

from codomyrmex.interpretability import SparseAutoencoder, train_sae, analyze_features

sae = train_sae(activations, d_features=256, n_steps=100, lambda_l1=1e-3)
analysis = analyze_features(sae, activations)

Exports

SparseAutoencoder, train_sae, analyze_features

Navigation