| { |
| "rows": [ |
| { |
| "metric": "l2", |
| "transformation": "identity_replay", |
| "adjusted_rand": 1.0, |
| "adjusted_rand_core_atoms_only": 1.0, |
| "n_modules_original": 3, |
| "n_modules_transformed": 3, |
| "exact_set_recovery": 1.0, |
| "mean_best_jaccard": 1.0, |
| "median_best_jaccard": 1.0, |
| "mean_best_jaccard_modules_le8": 1.0, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [ |
| 2, |
| 2, |
| 2 |
| ], |
| "per_module_best_jaccard": [ |
| 1.0, |
| 1.0, |
| 1.0 |
| ], |
| "note": "same data and rule, fresh run; upper bound on agreement achievable under any transformation" |
| }, |
| { |
| "metric": "l2", |
| "transformation": "positive_powers_of_two_scaling", |
| "gram_max_relative_change": 0.0, |
| "adjusted_rand": 1.0, |
| "adjusted_rand_core_atoms_only": 1.0, |
| "n_modules_original": 3, |
| "n_modules_transformed": 3, |
| "exact_set_recovery": 1.0, |
| "mean_best_jaccard": 1.0, |
| "median_best_jaccard": 1.0, |
| "mean_best_jaccard_modules_le8": 1.0, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [ |
| 2, |
| 2, |
| 2 |
| ], |
| "per_module_best_jaccard": [ |
| 1.0, |
| 1.0, |
| 1.0 |
| ], |
| "note": "c_j -> a_j c_j, v_j -> v_j/a_j applied to the cached coefficients, output weights and (Fisher) score projections; z_j preserved" |
| }, |
| { |
| "metric": "l2", |
| "transformation": "positive_half_split_32_atoms", |
| "cluster_count_mode": "free_C", |
| "adjusted_rand": -0.0036128001617856094, |
| "adjusted_rand_unsplit_atoms_only": -0.003494733857488641, |
| "adjusted_rand_core_atoms_only": 0.0, |
| "n_modules_original": 3, |
| "n_modules_transformed": 43, |
| "exact_set_recovery": 0.0, |
| "mean_best_jaccard": 0.0, |
| "median_best_jaccard": 0.0, |
| "mean_best_jaccard_modules_le8": 0.0, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [ |
| 2, |
| 2, |
| 2 |
| ], |
| "per_module_best_jaccard": [ |
| 0.0, |
| 0.0, |
| 0.0 |
| ], |
| "reference_C_original": 3, |
| "reference_C_transformed": 31, |
| "split_copies_co_assigned": 0.96875, |
| "split_copies_co_assigned_core_only": null, |
| "n_split_atoms_in_a_core_module": 0, |
| "split_atoms_in_a_core_module_originally": 0.0, |
| "affinity_max_relative_change_on_unsplit_atoms": { |
| "fcm": 0.0, |
| "cosine": 0.0 |
| }, |
| "collinear_collapse": null, |
| "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction." |
| }, |
| { |
| "metric": "l2", |
| "transformation": "positive_half_split_32_atoms", |
| "cluster_count_mode": "matched_C", |
| "adjusted_rand": -0.001934228314481383, |
| "adjusted_rand_unsplit_atoms_only": 0.0, |
| "adjusted_rand_core_atoms_only": 0.0, |
| "n_modules_original": 3, |
| "n_modules_transformed": 3, |
| "exact_set_recovery": 0.0, |
| "mean_best_jaccard": 0.0, |
| "median_best_jaccard": 0.0, |
| "mean_best_jaccard_modules_le8": 0.0, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [ |
| 2, |
| 2, |
| 2 |
| ], |
| "per_module_best_jaccard": [ |
| 0.0, |
| 0.0, |
| 0.0 |
| ], |
| "reference_C_original": 3, |
| "reference_C_transformed": 3, |
| "split_copies_co_assigned": 0.09375, |
| "split_copies_co_assigned_core_only": null, |
| "n_split_atoms_in_a_core_module": 0, |
| "split_atoms_in_a_core_module_originally": 0.0, |
| "affinity_max_relative_change_on_unsplit_atoms": { |
| "fcm": 0.0, |
| "cosine": 0.0 |
| }, |
| "collinear_collapse": null, |
| "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction." |
| }, |
| { |
| "metric": "l2", |
| "transformation": "positive_half_split_32_atoms", |
| "cluster_count_mode": "matched_C_collinear_collapsed", |
| "adjusted_rand": 1.0, |
| "adjusted_rand_unsplit_atoms_only": 1.0, |
| "adjusted_rand_core_atoms_only": 1.0, |
| "n_modules_original": 3, |
| "n_modules_transformed": 3, |
| "exact_set_recovery": 1.0, |
| "mean_best_jaccard": 1.0, |
| "median_best_jaccard": 1.0, |
| "mean_best_jaccard_modules_le8": 1.0, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [ |
| 2, |
| 2, |
| 2 |
| ], |
| "per_module_best_jaccard": [ |
| 1.0, |
| 1.0, |
| 1.0 |
| ], |
| "reference_C_original": 3, |
| "reference_C_transformed": 3, |
| "split_copies_co_assigned": 0.0, |
| "split_copies_co_assigned_core_only": null, |
| "n_split_atoms_in_a_core_module": 0, |
| "split_atoms_in_a_core_module_originally": 0.0, |
| "affinity_max_relative_change_on_unsplit_atoms": { |
| "fcm": 0.0, |
| "cosine": 0.0 |
| }, |
| "collinear_collapse": { |
| "atoms": 3104, |
| "merged_groups": 3072, |
| "merged_gram_max_relative_difference_from_original": 0.0, |
| "collinear_groups_in_original_layer": 3072, |
| "original_width": 3072 |
| }, |
| "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction." |
| }, |
| { |
| "metric": "l2", |
| "transformation": "baseline_raw_covariance_affinity_under_scaling", |
| "adjusted_rand": 0.8118744351996922, |
| "n_modules_original": 3, |
| "n_modules_transformed": 3, |
| "note": "naive co-activation clustering on W=|E[c_i c_j]| (no mass normalization); the same function-preserving scaling changes this affinity" |
| }, |
| { |
| "metric": "fisher", |
| "transformation": "identity_replay", |
| "adjusted_rand": 1.0, |
| "adjusted_rand_core_atoms_only": null, |
| "n_modules_original": 0, |
| "n_modules_transformed": 0, |
| "exact_set_recovery": 0.0, |
| "mean_best_jaccard": null, |
| "median_best_jaccard": null, |
| "mean_best_jaccard_modules_le8": null, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [], |
| "per_module_best_jaccard": [], |
| "note": "same data and rule, fresh run; upper bound on agreement achievable under any transformation" |
| }, |
| { |
| "metric": "fisher", |
| "transformation": "positive_powers_of_two_scaling", |
| "gram_max_relative_change": 0.0, |
| "adjusted_rand": 1.0, |
| "adjusted_rand_core_atoms_only": null, |
| "n_modules_original": 0, |
| "n_modules_transformed": 0, |
| "exact_set_recovery": 0.0, |
| "mean_best_jaccard": null, |
| "median_best_jaccard": null, |
| "mean_best_jaccard_modules_le8": null, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [], |
| "per_module_best_jaccard": [], |
| "note": "c_j -> a_j c_j, v_j -> v_j/a_j applied to the cached coefficients, output weights and (Fisher) score projections; z_j preserved" |
| }, |
| { |
| "metric": "fisher", |
| "transformation": "positive_half_split_32_atoms", |
| "cluster_count_mode": "free_C", |
| "adjusted_rand": 0.0, |
| "adjusted_rand_unsplit_atoms_only": 0.0, |
| "adjusted_rand_core_atoms_only": null, |
| "n_modules_original": 0, |
| "n_modules_transformed": 63, |
| "exact_set_recovery": 0.0, |
| "mean_best_jaccard": null, |
| "median_best_jaccard": null, |
| "mean_best_jaccard_modules_le8": null, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [], |
| "per_module_best_jaccard": [], |
| "reference_C_original": 2, |
| "reference_C_transformed": 30, |
| "split_copies_co_assigned": 0.96875, |
| "split_copies_co_assigned_core_only": null, |
| "n_split_atoms_in_a_core_module": 0, |
| "split_atoms_in_a_core_module_originally": 0.0, |
| "affinity_max_relative_change_on_unsplit_atoms": { |
| "fcm": 0.0, |
| "cosine": 0.0 |
| }, |
| "collinear_collapse": null, |
| "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction." |
| }, |
| { |
| "metric": "fisher", |
| "transformation": "positive_half_split_32_atoms", |
| "cluster_count_mode": "matched_C", |
| "adjusted_rand": 0.0, |
| "adjusted_rand_unsplit_atoms_only": 1.0, |
| "adjusted_rand_core_atoms_only": null, |
| "n_modules_original": 0, |
| "n_modules_transformed": 1, |
| "exact_set_recovery": 0.0, |
| "mean_best_jaccard": null, |
| "median_best_jaccard": null, |
| "mean_best_jaccard_modules_le8": null, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [], |
| "per_module_best_jaccard": [], |
| "reference_C_original": 2, |
| "reference_C_transformed": 2, |
| "split_copies_co_assigned": 0.03125, |
| "split_copies_co_assigned_core_only": null, |
| "n_split_atoms_in_a_core_module": 0, |
| "split_atoms_in_a_core_module_originally": 0.0, |
| "affinity_max_relative_change_on_unsplit_atoms": { |
| "fcm": 0.0, |
| "cosine": 0.0 |
| }, |
| "collinear_collapse": null, |
| "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction." |
| }, |
| { |
| "metric": "fisher", |
| "transformation": "positive_half_split_32_atoms", |
| "cluster_count_mode": "matched_C_collinear_collapsed", |
| "adjusted_rand": 1.0, |
| "adjusted_rand_unsplit_atoms_only": 1.0, |
| "adjusted_rand_core_atoms_only": null, |
| "n_modules_original": 0, |
| "n_modules_transformed": 0, |
| "exact_set_recovery": 0.0, |
| "mean_best_jaccard": null, |
| "median_best_jaccard": null, |
| "mean_best_jaccard_modules_le8": null, |
| "mean_best_jaccard_modules_gt8": null, |
| "module_sizes": [], |
| "per_module_best_jaccard": [], |
| "reference_C_original": 2, |
| "reference_C_transformed": 2, |
| "split_copies_co_assigned": 0.0, |
| "split_copies_co_assigned_core_only": null, |
| "n_split_atoms_in_a_core_module": 0, |
| "split_atoms_in_a_core_module_originally": 0.0, |
| "affinity_max_relative_change_on_unsplit_atoms": { |
| "fcm": 0.0, |
| "cosine": 0.0 |
| }, |
| "collinear_collapse": { |
| "atoms": 3104, |
| "merged_groups": 3072, |
| "merged_gram_max_relative_difference_from_original": 0.0, |
| "collinear_groups_in_original_layer": 3072, |
| "original_width": 3072 |
| }, |
| "note": "duplicate atoms with halved output weights; ARI lifts the original labels through the duplication index. Each split pair is a perfectly coupled 2-clique that adds a near-zero normalized-Laplacian eigenvalue, so the eigengap rule reacts to the duplicates; matched_C removes that reaction." |
| }, |
| { |
| "metric": "fisher", |
| "transformation": "baseline_raw_covariance_affinity_under_scaling", |
| "adjusted_rand": 0.8118744351996922, |
| "n_modules_original": 3, |
| "n_modules_transformed": 3, |
| "note": "naive co-activation clustering on W=|E[c_i c_j]| (no mass normalization); the same function-preserving scaling changes this affinity" |
| } |
| ], |
| "settings": { |
| "root": "results/comparison/random_gpt2", |
| "out": "results/modules/random_gpt2", |
| "metrics": [ |
| "l2", |
| "fisher" |
| ], |
| "bootstraps": 12, |
| "split_atoms": 32, |
| "seed": 17, |
| "affinity": "square", |
| "knn": 0, |
| "normalization": "fcm", |
| "cluster_rule": "reference", |
| "clusters": null, |
| "min_clusters": 2, |
| "max_clusters": 32, |
| "threshold": 0.8, |
| "min_size": 2, |
| "max_module_fraction": 0.5, |
| "restarts": 5 |
| }, |
| "seconds": 29.21317982673645, |
| "scope": "partition identity only; semantic and causal changes under reparameterization are not remeasured here", |
| "ari_convention": "the rest module is treated as one additional label, so ARI scores the whole assignment including which neurons are declared unstructured" |
| } |
