omega-activation-feature-factorization

Use when real activation matrices are available and interpretable factorization is useful. Runs bounded non-negative matrix factorization and reports reconstruction/sparsity without claiming universal monosemantic features.