diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/code_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x16/l0_128/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..854b8ae336881077d46322bd7310278687f599ff --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 303.350046692607, + "recon_r2": 0.9945499165194519, + "l1_per_token": 235.00570939926905 + }, + "seed1": { + "l0": 308.26556420233464, + "recon_r2": 0.9947738428369166, + "l1_per_token": 237.94410117398962 + }, + "seed2": { + "l0": 303.32809208819714, + "recon_r2": 0.994642080225655, + "l1_per_token": 236.53932976929894 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.3081740635698948, + "mcc_abs": 0.308183846174609, + "act_iou": 0.06158082759211115, + "act_iou_n_pairs": 12288, + "input_r2": 0.9944915843011223, + "matched_mcc_signed": { + "0.6": 0.486311554204309, + "0.7": 0.47742297952236057, + "0.8": 0.4724732855685592, + "0.9": 0.499113829069671 + }, + "matched_mcc_abs": { + "0.6": 0.48502962515759485, + "0.7": 0.47719360709953557, + "0.8": 0.4724564351388078, + "0.9": 0.49935382096494135 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.30419426695625673, + "mcc_abs": 0.30420468250052374, + "act_iou": 0.06069953939073782, + "act_iou_n_pairs": 12288, + "input_r2": 0.994404892243357, + "matched_mcc_signed": { + "0.6": 0.48488945895927105, + "0.7": 0.47440591114644926, + "0.8": 0.4711649853133228, + "0.9": 0.49944654239146674 + }, + "matched_mcc_abs": { + "0.6": 0.4840749925413885, + "0.7": 0.4744411480351126, + "0.8": 0.46963779513166715, + "0.9": 0.4997708880092602 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.3070745853473141, + "mcc_abs": 0.3070853782622768, + "act_iou": 0.06065747285174358, + "act_iou_n_pairs": 12288, + "input_r2": 0.994486634330592, + "matched_mcc_signed": { + "0.6": 0.4903236837017785, + "0.7": 0.48064124130373964, + "0.8": 0.4743521280493735, + "0.9": 0.4958878012745798 + }, + "matched_mcc_abs": { + "0.6": 0.48972885496687907, + "0.7": 0.4805835291198603, + "0.8": 0.4730622072347284, + "0.9": 0.49558014937754513 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 304.9812343277129, + "l0_std": 2.8443342885749576, + "recon_r2_mean": 0.9946552798606746, + "recon_r2_std": 0.00011254519841739937, + "l1_per_token_mean": 236.49638011418588, + "l1_per_token_std": 1.4696666493093344, + "input_r2_mean": 0.9944610369583572, + "input_r2_std": 4.868569932631574e-05, + "mcc_signed_mean": 0.30648097195782187, + "mcc_signed_std": 0.0020552318393844463, + "mcc_abs_mean": 0.30649130231246985, + "mcc_abs_std": 0.0020550256839986587, + "matched_mcc_signed_per_threshold": { + "0.6": 0.4871748989551195, + "0.7": 0.4774900439908498, + "0.8": 0.4726634663104185, + "0.9": 0.49814939091190585 + }, + "matched_mcc_signed_band": 0.48386945004207343, + "matched_mcc_abs_per_threshold": { + "0.6": 0.48627782422195415, + "0.7": 0.47740609475150286, + "0.8": 0.4717188125017344, + "0.9": 0.49823495278391555 + }, + "matched_mcc_abs_band": 0.48340942106477675, + "act_iou_mean": 0.060979279944864184, + "act_iou_std": 0.0005213799740659474, + "act_iou_n_pairs": 12288.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/dict_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x16/l0_128/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..8cd42332d4b9868b8200c9463fc9dca191f3c031 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/dict_metrics_step_final.json @@ -0,0 +1,545 @@ +{ + "schema": 2, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8607, + "0.6": 7559, + "0.7": 6051, + "0.8": 3854, + "0.9": 1228 + }, + "n_dup_clusters_signed": { + "0.5": 2371, + "0.6": 2294, + "0.7": 2002, + "0.8": 1407, + "0.9": 519 + }, + "largest_dup_signed": { + "0.5": 367, + "0.6": 236, + "0.7": 134, + "0.8": 29, + "0.9": 10 + }, + "babel_mu1_signed": 0.6670389771461487, + "n_dup_abs": { + "0.5": 8643, + "0.6": 7607, + "0.7": 6081, + "0.8": 3874, + "0.9": 1228 + }, + "n_dup_clusters_abs": { + "0.5": 2386, + "0.6": 2311, + "0.7": 2013, + "0.8": 1414, + "0.9": 517 + }, + "largest_dup_abs": { + "0.5": 372, + "0.6": 295, + "0.7": 114, + "0.8": 29, + "0.9": 10 + }, + "babel_mu1_abs": 0.6695666313171387 + }, + "seed1": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8532, + "0.6": 7493, + "0.7": 6077, + "0.8": 3854, + "0.9": 1170 + }, + "n_dup_clusters_signed": { + "0.5": 2322, + "0.6": 2234, + "0.7": 2022, + "0.8": 1405, + "0.9": 494 + }, + "largest_dup_signed": { + "0.5": 439, + "0.6": 325, + "0.7": 145, + "0.8": 50, + "0.9": 10 + }, + "babel_mu1_signed": 0.6654933094978333, + "n_dup_abs": { + "0.5": 8570, + "0.6": 7525, + "0.7": 6107, + "0.8": 3874, + "0.9": 1172 + }, + "n_dup_clusters_abs": { + "0.5": 2326, + "0.6": 2245, + "0.7": 2036, + "0.8": 1414, + "0.9": 494 + }, + "largest_dup_abs": { + "0.5": 235, + "0.6": 173, + "0.7": 147, + "0.8": 50, + "0.9": 10 + }, + "babel_mu1_abs": 0.667901337146759 + }, + "seed2": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8456, + "0.6": 7442, + "0.7": 6006, + "0.8": 3793, + "0.9": 1148 + }, + "n_dup_clusters_signed": { + "0.5": 2327, + "0.6": 2258, + "0.7": 1995, + "0.8": 1400, + "0.9": 485 + }, + "largest_dup_signed": { + "0.5": 355, + "0.6": 144, + "0.7": 77, + "0.8": 36, + "0.9": 10 + }, + "babel_mu1_signed": 0.6603737473487854, + "n_dup_abs": { + "0.5": 8512, + "0.6": 7487, + "0.7": 6050, + "0.8": 3818, + "0.9": 1152 + }, + "n_dup_clusters_abs": { + "0.5": 2346, + "0.6": 2272, + "0.7": 2011, + "0.8": 1409, + "0.9": 485 + }, + "largest_dup_abs": { + "0.5": 338, + "0.6": 286, + "0.7": 88, + "0.8": 36, + "0.9": 10 + }, + "babel_mu1_abs": 0.6630961298942566 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.7817156910896301, + "dcs_abs": 0.7834139466285706, + "frac_sign_flipped": 0.039388020833333336, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.7236672043800354, + "0.6": 0.7293717265129089, + "0.7": 0.7402892112731934, + "0.8": 0.7550824880599976, + "0.9": 0.7738751173019409 + }, + "unique_frac_signed": { + "0.5": 0.4935709635416667, + "0.6": 0.57177734375, + "0.7": 0.6702473958333334, + "0.8": 0.80078125, + "0.9": 0.9436442057291666 + }, + "dedup_dcs_signed": { + "0.5": 0.3571811193493583, + "0.6": 0.4170382283919025, + "0.7": 0.4961769160193702, + "0.8": 0.6046558986417949, + "0.9": 0.7302627703999557 + }, + "matched_frac_signed": { + "0.5": 0.3589680989583333, + "0.6": 0.4026692708333333, + "0.7": 0.4476725260416667, + "0.8": 0.482421875, + "0.9": 0.443359375 + }, + "matched_cos_abs": { + "0.5": 0.7263489365577698, + "0.6": 0.732623279094696, + "0.7": 0.7427051663398743, + "0.8": 0.7570679783821106, + "0.9": 0.7756255269050598 + }, + "unique_frac_abs": { + "0.5": 0.4913330078125, + "0.6": 0.5696614583333334, + "0.7": 0.6688232421875, + "0.8": 0.7998046875, + "0.9": 0.9434814453125 + }, + "dedup_dcs_abs": { + "0.5": 0.35687920772033976, + "0.6": 0.41734724557803327, + "0.7": 0.4967384773408412, + "0.8": 0.6055065178661607, + "0.9": 0.7317882931456552 + }, + "matched_frac_abs": { + "0.5": 0.3590494791666667, + "0.6": 0.4029947916666667, + "0.7": 0.447998046875, + "0.8": 0.4825032552083333, + "0.9": 0.443359375 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.7770745754241943, + "dcs_abs": 0.7787559628486633, + "frac_sign_flipped": 0.043701171875, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.7226144671440125, + "0.6": 0.7275781631469727, + "0.7": 0.7358648180961609, + "0.8": 0.7517416477203369, + "0.9": 0.7691615223884583 + }, + "unique_frac_signed": { + "0.5": 0.4968668619791667, + "0.6": 0.5748291015625, + "0.7": 0.6720377604166666, + "0.8": 0.8030598958333334, + "0.9": 0.9441731770833334 + }, + "dedup_dcs_signed": { + "0.5": 0.3590431827105931, + "0.6": 0.41823310183826834, + "0.7": 0.49452894432276173, + "0.8": 0.6036935693118721, + "0.9": 0.726221678283764 + }, + "matched_frac_signed": { + "0.5": 0.3600260416666667, + "0.6": 0.4007161458333333, + "0.7": 0.4436848958333333, + "0.8": 0.4747721354166667, + "0.9": 0.4336751302083333 + }, + "matched_cos_abs": { + "0.5": 0.7251178026199341, + "0.6": 0.7300339341163635, + "0.7": 0.737865686416626, + "0.8": 0.7535355687141418, + "0.9": 0.7708365321159363 + }, + "unique_frac_abs": { + "0.5": 0.4945068359375, + "0.6": 0.5723063151041666, + "0.7": 0.6701253255208334, + "0.8": 0.8018798828125, + "0.9": 0.9439290364583334 + }, + "dedup_dcs_abs": { + "0.5": 0.35857571025553625, + "0.6": 0.417803030735134, + "0.7": 0.4944624833005946, + "0.8": 0.6042450135355466, + "0.9": 0.7276149850270789 + }, + "matched_frac_abs": { + "0.5": 0.3597005208333333, + "0.6": 0.4002278645833333, + "0.7": 0.443115234375, + "0.8": 0.475830078125, + "0.9": 0.4337565104166667 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.7794721722602844, + "dcs_abs": 0.781230628490448, + "frac_sign_flipped": 0.042073567708333336, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.722134530544281, + "0.6": 0.7282166481018066, + "0.7": 0.7373058795928955, + "0.8": 0.7540035247802734, + "0.9": 0.7709962725639343 + }, + "unique_frac_signed": { + "0.5": 0.4979248046875, + "0.6": 0.5750732421875, + "0.7": 0.6717936197916666, + "0.8": 0.802978515625, + "0.9": 0.9455159505208334 + }, + "dedup_dcs_signed": { + "0.5": 0.3595686950793606, + "0.6": 0.4187779088388197, + "0.7": 0.49531738574538997, + "0.8": 0.6054486311040819, + "0.9": 0.7289892735013079 + }, + "matched_frac_signed": { + "0.5": 0.3595377604166667, + "0.6": 0.3990885416666667, + "0.7": 0.4423828125, + "0.8": 0.47998046875, + "0.9": 0.4520670572916667 + }, + "matched_cos_abs": { + "0.5": 0.7249367237091064, + "0.6": 0.7299293875694275, + "0.7": 0.7390979528427124, + "0.8": 0.7557072639465332, + "0.9": 0.7726919054985046 + }, + "unique_frac_abs": { + "0.5": 0.4950358072916667, + "0.6": 0.5729573567708334, + "0.7": 0.6700032552083334, + "0.8": 0.8018798828125, + "0.9": 0.9452718098958334 + }, + "dedup_dcs_abs": { + "0.5": 0.3588696362567134, + "0.6": 0.41821841253113234, + "0.7": 0.4951980343224326, + "0.8": 0.605986452254001, + "0.9": 0.7304038760024317 + }, + "matched_frac_abs": { + "0.5": 0.359130859375, + "0.6": 0.3982747395833333, + "0.7": 0.442138671875, + "0.8": 0.48046875, + "0.9": 0.452392578125 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 0, + "n_dense_std": 0.0, + "l0_mean": null, + "l0_std": null, + "recon_r2_mean": null, + "recon_r2_std": null, + "l1_per_token_mean": null, + "l1_per_token_std": null, + "babel_mu1_signed_mean": 0.6643020113309225, + "babel_mu1_signed_std": 0.0034886552645365565, + "n_dup_signed_per_threshold": { + "0.5": 8531.666666666666, + "0.6": 7498.0, + "0.7": 6044.666666666667, + "0.8": 3833.6666666666665, + "0.9": 1182.0 + }, + "n_dup_signed_band": 4639.583333333334, + "n_dup_signed_band_std": 38.65418261111381, + "n_dup_clusters_signed_per_threshold": { + "0.5": 2340.0, + "0.6": 2262.0, + "0.7": 2006.3333333333333, + "0.8": 1404.0, + "0.9": 499.3333333333333 + }, + "n_dup_clusters_signed_band": 1542.9166666666665, + "n_dup_clusters_signed_band_std": 11.10273990208423, + "largest_dup_signed_per_threshold": { + "0.5": 387.0, + "0.6": 235.0, + "0.7": 118.66666666666667, + "0.8": 38.333333333333336, + "0.9": 10.0 + }, + "largest_dup_signed_band": 100.5, + "largest_dup_signed_band_std": 32.90991491936739, + "dcs_signed_mean": 0.779420812924703, + "dcs_signed_std": 0.0023209840566022644, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.35859766571310403, + "0.6": 0.41801641302299686, + "0.7": 0.49534108202917393, + "0.8": 0.6045993663525829, + "0.9": 0.7284912407283425 + }, + "dedup_dcs_signed_band": 0.561612025533274, + "dedup_dcs_signed_band_std": 0.000817928943294489, + "matched_cos_signed_per_threshold": { + "0.5": 0.722805400689443, + "0.6": 0.7283888459205627, + "0.7": 0.7378199696540833, + "0.8": 0.7536092201868693, + "0.9": 0.7713443040847778 + }, + "matched_cos_signed_band": 0.7477905849615732, + "matched_cos_signed_band_std": 0.001789422150281412, + "unique_frac_signed_per_threshold": { + "0.5": 0.49612087673611116, + "0.6": 0.5738932291666666, + "0.7": 0.6713595920138888, + "0.8": 0.8022732204861112, + "0.9": 0.9444444444444445 + }, + "unique_frac_signed_band": 0.7479926215277778, + "unique_frac_signed_band_std": 0.0012055337535602764, + "matched_frac_signed_per_threshold": { + "0.5": 0.3595106336805556, + "0.6": 0.40082465277777773, + "0.7": 0.444580078125, + "0.8": 0.47905815972222227, + "0.9": 0.4430338541666667 + }, + "matched_frac_signed_band": 0.4418741861979167, + "matched_frac_signed_band_std": 0.003188141718741717, + "babel_mu1_abs_mean": 0.6668546994527181, + "babel_mu1_abs_std": 0.0033598266343399697, + "n_dup_abs_per_threshold": { + "0.5": 8575.0, + "0.6": 7539.666666666667, + "0.7": 6079.333333333333, + "0.8": 3855.3333333333335, + "0.9": 1184.0 + }, + "n_dup_abs_band": 4664.583333333333, + "n_dup_abs_band_std": 35.63033585771166, + "n_dup_clusters_abs_per_threshold": { + "0.5": 2352.6666666666665, + "0.6": 2276.0, + "0.7": 2020.0, + "0.8": 1412.3333333333333, + "0.9": 498.6666666666667 + }, + "n_dup_clusters_abs_band": 1551.75, + "n_dup_clusters_abs_band_std": 10.5, + "largest_dup_abs_per_threshold": { + "0.5": 315.0, + "0.6": 251.33333333333334, + "0.7": 116.33333333333333, + "0.8": 38.333333333333336, + "0.9": 10.0 + }, + "largest_dup_abs_band": 104.0, + "largest_dup_abs_band_std": 8.54400374531753, + "dcs_abs_mean": 0.7811335126558939, + "dcs_abs_std": 0.0023305099951428116, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.3581081847441965, + "0.6": 0.4177895629480999, + "0.7": 0.49546633165462284, + "0.8": 0.6052459945519028, + "0.9": 0.7299357180583886 + }, + "dedup_dcs_abs_band": 0.5621094018032535, + "dedup_dcs_abs_band_std": 0.0009540964043521437, + "matched_cos_abs_per_threshold": { + "0.5": 0.7254678209622701, + "0.6": 0.7308622002601624, + "0.7": 0.7398896018664042, + "0.8": 0.7554369370142618, + "0.9": 0.7730513215065002 + }, + "matched_cos_abs_band": 0.7498100151618321, + "matched_cos_abs_band_std": 0.0020075506852886187, + "unique_frac_abs_per_threshold": { + "0.5": 0.4936252170138889, + "0.6": 0.5716417100694445, + "0.7": 0.669650607638889, + "0.8": 0.8011881510416666, + "0.9": 0.9442274305555557 + }, + "unique_frac_abs_band": 0.746676974826389, + "unique_frac_abs_band_std": 0.0010942126864736965, + "matched_frac_abs_per_threshold": { + "0.5": 0.3592936197916667, + "0.6": 0.4004991319444444, + "0.7": 0.4444173177083333, + "0.8": 0.4796006944444444, + "0.9": 0.44316948784722227 + }, + "matched_frac_abs_band": 0.4419216579861111, + "matched_frac_abs_band_std": 0.003226171902808814, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/seed0/config.json b/vision/dinov2-b_layer10/mp/x16/l0_128/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..582c8dce741510337aae6bc33e8db9b021f25bc1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/seed0/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..8362eb3e73c18d7ca879ba894c4f1d98d685af85 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1298, + "r2_score": 0.9648, + "l0_per_token_0": 0.0029, + "l0_per_token_0.001": 0.0029, + "l0_per_token_0.01": 0.0029, + "l0_per_token_0.05": 0.0029, + "l0_per_token_0.1": 0.0029, + "l0_per_token_0.5": 0.0029, + "l0_per_token_1.0": 0.0026, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 96.7825 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/seed1/config.json b/vision/dinov2-b_layer10/mp/x16/l0_128/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8465153a9e9a1f2031f87935b38abdc81b8f3232 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/seed1/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..53e35fd12396d401b0aad7c91327037b5e6a76a8 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1258, + "r2_score": 0.9659, + "l0_per_token_0": 0.0031, + "l0_per_token_0.001": 0.0031, + "l0_per_token_0.01": 0.0031, + "l0_per_token_0.05": 0.0031, + "l0_per_token_0.1": 0.0031, + "l0_per_token_0.5": 0.0031, + "l0_per_token_1.0": 0.0027, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 100.9962 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/seed2/config.json b/vision/dinov2-b_layer10/mp/x16/l0_128/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..859d3bae315c9ef5dc1a4ece37711d5e66c114f1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/seed2/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..e904d1368bf2bf3804acc8a9aa983626ad162ba9 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1287, + "r2_score": 0.9651, + "l0_per_token_0": 0.003, + "l0_per_token_0.001": 0.003, + "l0_per_token_0.01": 0.003, + "l0_per_token_0.05": 0.003, + "l0_per_token_0.1": 0.003, + "l0_per_token_0.5": 0.003, + "l0_per_token_1.0": 0.0026, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 98.9773 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/code_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x16/l0_32/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..854b8ae336881077d46322bd7310278687f599ff --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 303.350046692607, + "recon_r2": 0.9945499165194519, + "l1_per_token": 235.00570939926905 + }, + "seed1": { + "l0": 308.26556420233464, + "recon_r2": 0.9947738428369166, + "l1_per_token": 237.94410117398962 + }, + "seed2": { + "l0": 303.32809208819714, + "recon_r2": 0.994642080225655, + "l1_per_token": 236.53932976929894 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.3081740635698948, + "mcc_abs": 0.308183846174609, + "act_iou": 0.06158082759211115, + "act_iou_n_pairs": 12288, + "input_r2": 0.9944915843011223, + "matched_mcc_signed": { + "0.6": 0.486311554204309, + "0.7": 0.47742297952236057, + "0.8": 0.4724732855685592, + "0.9": 0.499113829069671 + }, + "matched_mcc_abs": { + "0.6": 0.48502962515759485, + "0.7": 0.47719360709953557, + "0.8": 0.4724564351388078, + "0.9": 0.49935382096494135 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.30419426695625673, + "mcc_abs": 0.30420468250052374, + "act_iou": 0.06069953939073782, + "act_iou_n_pairs": 12288, + "input_r2": 0.994404892243357, + "matched_mcc_signed": { + "0.6": 0.48488945895927105, + "0.7": 0.47440591114644926, + "0.8": 0.4711649853133228, + "0.9": 0.49944654239146674 + }, + "matched_mcc_abs": { + "0.6": 0.4840749925413885, + "0.7": 0.4744411480351126, + "0.8": 0.46963779513166715, + "0.9": 0.4997708880092602 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.3070745853473141, + "mcc_abs": 0.3070853782622768, + "act_iou": 0.06065747285174358, + "act_iou_n_pairs": 12288, + "input_r2": 0.994486634330592, + "matched_mcc_signed": { + "0.6": 0.4903236837017785, + "0.7": 0.48064124130373964, + "0.8": 0.4743521280493735, + "0.9": 0.4958878012745798 + }, + "matched_mcc_abs": { + "0.6": 0.48972885496687907, + "0.7": 0.4805835291198603, + "0.8": 0.4730622072347284, + "0.9": 0.49558014937754513 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 304.9812343277129, + "l0_std": 2.8443342885749576, + "recon_r2_mean": 0.9946552798606746, + "recon_r2_std": 0.00011254519841739937, + "l1_per_token_mean": 236.49638011418588, + "l1_per_token_std": 1.4696666493093344, + "input_r2_mean": 0.9944610369583572, + "input_r2_std": 4.868569932631574e-05, + "mcc_signed_mean": 0.30648097195782187, + "mcc_signed_std": 0.0020552318393844463, + "mcc_abs_mean": 0.30649130231246985, + "mcc_abs_std": 0.0020550256839986587, + "matched_mcc_signed_per_threshold": { + "0.6": 0.4871748989551195, + "0.7": 0.4774900439908498, + "0.8": 0.4726634663104185, + "0.9": 0.49814939091190585 + }, + "matched_mcc_signed_band": 0.48386945004207343, + "matched_mcc_abs_per_threshold": { + "0.6": 0.48627782422195415, + "0.7": 0.47740609475150286, + "0.8": 0.4717188125017344, + "0.9": 0.49823495278391555 + }, + "matched_mcc_abs_band": 0.48340942106477675, + "act_iou_mean": 0.060979279944864184, + "act_iou_std": 0.0005213799740659474, + "act_iou_n_pairs": 12288.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/dict_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x16/l0_32/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..8cd42332d4b9868b8200c9463fc9dca191f3c031 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/dict_metrics_step_final.json @@ -0,0 +1,545 @@ +{ + "schema": 2, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8607, + "0.6": 7559, + "0.7": 6051, + "0.8": 3854, + "0.9": 1228 + }, + "n_dup_clusters_signed": { + "0.5": 2371, + "0.6": 2294, + "0.7": 2002, + "0.8": 1407, + "0.9": 519 + }, + "largest_dup_signed": { + "0.5": 367, + "0.6": 236, + "0.7": 134, + "0.8": 29, + "0.9": 10 + }, + "babel_mu1_signed": 0.6670389771461487, + "n_dup_abs": { + "0.5": 8643, + "0.6": 7607, + "0.7": 6081, + "0.8": 3874, + "0.9": 1228 + }, + "n_dup_clusters_abs": { + "0.5": 2386, + "0.6": 2311, + "0.7": 2013, + "0.8": 1414, + "0.9": 517 + }, + "largest_dup_abs": { + "0.5": 372, + "0.6": 295, + "0.7": 114, + "0.8": 29, + "0.9": 10 + }, + "babel_mu1_abs": 0.6695666313171387 + }, + "seed1": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8532, + "0.6": 7493, + "0.7": 6077, + "0.8": 3854, + "0.9": 1170 + }, + "n_dup_clusters_signed": { + "0.5": 2322, + "0.6": 2234, + "0.7": 2022, + "0.8": 1405, + "0.9": 494 + }, + "largest_dup_signed": { + "0.5": 439, + "0.6": 325, + "0.7": 145, + "0.8": 50, + "0.9": 10 + }, + "babel_mu1_signed": 0.6654933094978333, + "n_dup_abs": { + "0.5": 8570, + "0.6": 7525, + "0.7": 6107, + "0.8": 3874, + "0.9": 1172 + }, + "n_dup_clusters_abs": { + "0.5": 2326, + "0.6": 2245, + "0.7": 2036, + "0.8": 1414, + "0.9": 494 + }, + "largest_dup_abs": { + "0.5": 235, + "0.6": 173, + "0.7": 147, + "0.8": 50, + "0.9": 10 + }, + "babel_mu1_abs": 0.667901337146759 + }, + "seed2": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8456, + "0.6": 7442, + "0.7": 6006, + "0.8": 3793, + "0.9": 1148 + }, + "n_dup_clusters_signed": { + "0.5": 2327, + "0.6": 2258, + "0.7": 1995, + "0.8": 1400, + "0.9": 485 + }, + "largest_dup_signed": { + "0.5": 355, + "0.6": 144, + "0.7": 77, + "0.8": 36, + "0.9": 10 + }, + "babel_mu1_signed": 0.6603737473487854, + "n_dup_abs": { + "0.5": 8512, + "0.6": 7487, + "0.7": 6050, + "0.8": 3818, + "0.9": 1152 + }, + "n_dup_clusters_abs": { + "0.5": 2346, + "0.6": 2272, + "0.7": 2011, + "0.8": 1409, + "0.9": 485 + }, + "largest_dup_abs": { + "0.5": 338, + "0.6": 286, + "0.7": 88, + "0.8": 36, + "0.9": 10 + }, + "babel_mu1_abs": 0.6630961298942566 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.7817156910896301, + "dcs_abs": 0.7834139466285706, + "frac_sign_flipped": 0.039388020833333336, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.7236672043800354, + "0.6": 0.7293717265129089, + "0.7": 0.7402892112731934, + "0.8": 0.7550824880599976, + "0.9": 0.7738751173019409 + }, + "unique_frac_signed": { + "0.5": 0.4935709635416667, + "0.6": 0.57177734375, + "0.7": 0.6702473958333334, + "0.8": 0.80078125, + "0.9": 0.9436442057291666 + }, + "dedup_dcs_signed": { + "0.5": 0.3571811193493583, + "0.6": 0.4170382283919025, + "0.7": 0.4961769160193702, + "0.8": 0.6046558986417949, + "0.9": 0.7302627703999557 + }, + "matched_frac_signed": { + "0.5": 0.3589680989583333, + "0.6": 0.4026692708333333, + "0.7": 0.4476725260416667, + "0.8": 0.482421875, + "0.9": 0.443359375 + }, + "matched_cos_abs": { + "0.5": 0.7263489365577698, + "0.6": 0.732623279094696, + "0.7": 0.7427051663398743, + "0.8": 0.7570679783821106, + "0.9": 0.7756255269050598 + }, + "unique_frac_abs": { + "0.5": 0.4913330078125, + "0.6": 0.5696614583333334, + "0.7": 0.6688232421875, + "0.8": 0.7998046875, + "0.9": 0.9434814453125 + }, + "dedup_dcs_abs": { + "0.5": 0.35687920772033976, + "0.6": 0.41734724557803327, + "0.7": 0.4967384773408412, + "0.8": 0.6055065178661607, + "0.9": 0.7317882931456552 + }, + "matched_frac_abs": { + "0.5": 0.3590494791666667, + "0.6": 0.4029947916666667, + "0.7": 0.447998046875, + "0.8": 0.4825032552083333, + "0.9": 0.443359375 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.7770745754241943, + "dcs_abs": 0.7787559628486633, + "frac_sign_flipped": 0.043701171875, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.7226144671440125, + "0.6": 0.7275781631469727, + "0.7": 0.7358648180961609, + "0.8": 0.7517416477203369, + "0.9": 0.7691615223884583 + }, + "unique_frac_signed": { + "0.5": 0.4968668619791667, + "0.6": 0.5748291015625, + "0.7": 0.6720377604166666, + "0.8": 0.8030598958333334, + "0.9": 0.9441731770833334 + }, + "dedup_dcs_signed": { + "0.5": 0.3590431827105931, + "0.6": 0.41823310183826834, + "0.7": 0.49452894432276173, + "0.8": 0.6036935693118721, + "0.9": 0.726221678283764 + }, + "matched_frac_signed": { + "0.5": 0.3600260416666667, + "0.6": 0.4007161458333333, + "0.7": 0.4436848958333333, + "0.8": 0.4747721354166667, + "0.9": 0.4336751302083333 + }, + "matched_cos_abs": { + "0.5": 0.7251178026199341, + "0.6": 0.7300339341163635, + "0.7": 0.737865686416626, + "0.8": 0.7535355687141418, + "0.9": 0.7708365321159363 + }, + "unique_frac_abs": { + "0.5": 0.4945068359375, + "0.6": 0.5723063151041666, + "0.7": 0.6701253255208334, + "0.8": 0.8018798828125, + "0.9": 0.9439290364583334 + }, + "dedup_dcs_abs": { + "0.5": 0.35857571025553625, + "0.6": 0.417803030735134, + "0.7": 0.4944624833005946, + "0.8": 0.6042450135355466, + "0.9": 0.7276149850270789 + }, + "matched_frac_abs": { + "0.5": 0.3597005208333333, + "0.6": 0.4002278645833333, + "0.7": 0.443115234375, + "0.8": 0.475830078125, + "0.9": 0.4337565104166667 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.7794721722602844, + "dcs_abs": 0.781230628490448, + "frac_sign_flipped": 0.042073567708333336, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.722134530544281, + "0.6": 0.7282166481018066, + "0.7": 0.7373058795928955, + "0.8": 0.7540035247802734, + "0.9": 0.7709962725639343 + }, + "unique_frac_signed": { + "0.5": 0.4979248046875, + "0.6": 0.5750732421875, + "0.7": 0.6717936197916666, + "0.8": 0.802978515625, + "0.9": 0.9455159505208334 + }, + "dedup_dcs_signed": { + "0.5": 0.3595686950793606, + "0.6": 0.4187779088388197, + "0.7": 0.49531738574538997, + "0.8": 0.6054486311040819, + "0.9": 0.7289892735013079 + }, + "matched_frac_signed": { + "0.5": 0.3595377604166667, + "0.6": 0.3990885416666667, + "0.7": 0.4423828125, + "0.8": 0.47998046875, + "0.9": 0.4520670572916667 + }, + "matched_cos_abs": { + "0.5": 0.7249367237091064, + "0.6": 0.7299293875694275, + "0.7": 0.7390979528427124, + "0.8": 0.7557072639465332, + "0.9": 0.7726919054985046 + }, + "unique_frac_abs": { + "0.5": 0.4950358072916667, + "0.6": 0.5729573567708334, + "0.7": 0.6700032552083334, + "0.8": 0.8018798828125, + "0.9": 0.9452718098958334 + }, + "dedup_dcs_abs": { + "0.5": 0.3588696362567134, + "0.6": 0.41821841253113234, + "0.7": 0.4951980343224326, + "0.8": 0.605986452254001, + "0.9": 0.7304038760024317 + }, + "matched_frac_abs": { + "0.5": 0.359130859375, + "0.6": 0.3982747395833333, + "0.7": 0.442138671875, + "0.8": 0.48046875, + "0.9": 0.452392578125 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 0, + "n_dense_std": 0.0, + "l0_mean": null, + "l0_std": null, + "recon_r2_mean": null, + "recon_r2_std": null, + "l1_per_token_mean": null, + "l1_per_token_std": null, + "babel_mu1_signed_mean": 0.6643020113309225, + "babel_mu1_signed_std": 0.0034886552645365565, + "n_dup_signed_per_threshold": { + "0.5": 8531.666666666666, + "0.6": 7498.0, + "0.7": 6044.666666666667, + "0.8": 3833.6666666666665, + "0.9": 1182.0 + }, + "n_dup_signed_band": 4639.583333333334, + "n_dup_signed_band_std": 38.65418261111381, + "n_dup_clusters_signed_per_threshold": { + "0.5": 2340.0, + "0.6": 2262.0, + "0.7": 2006.3333333333333, + "0.8": 1404.0, + "0.9": 499.3333333333333 + }, + "n_dup_clusters_signed_band": 1542.9166666666665, + "n_dup_clusters_signed_band_std": 11.10273990208423, + "largest_dup_signed_per_threshold": { + "0.5": 387.0, + "0.6": 235.0, + "0.7": 118.66666666666667, + "0.8": 38.333333333333336, + "0.9": 10.0 + }, + "largest_dup_signed_band": 100.5, + "largest_dup_signed_band_std": 32.90991491936739, + "dcs_signed_mean": 0.779420812924703, + "dcs_signed_std": 0.0023209840566022644, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.35859766571310403, + "0.6": 0.41801641302299686, + "0.7": 0.49534108202917393, + "0.8": 0.6045993663525829, + "0.9": 0.7284912407283425 + }, + "dedup_dcs_signed_band": 0.561612025533274, + "dedup_dcs_signed_band_std": 0.000817928943294489, + "matched_cos_signed_per_threshold": { + "0.5": 0.722805400689443, + "0.6": 0.7283888459205627, + "0.7": 0.7378199696540833, + "0.8": 0.7536092201868693, + "0.9": 0.7713443040847778 + }, + "matched_cos_signed_band": 0.7477905849615732, + "matched_cos_signed_band_std": 0.001789422150281412, + "unique_frac_signed_per_threshold": { + "0.5": 0.49612087673611116, + "0.6": 0.5738932291666666, + "0.7": 0.6713595920138888, + "0.8": 0.8022732204861112, + "0.9": 0.9444444444444445 + }, + "unique_frac_signed_band": 0.7479926215277778, + "unique_frac_signed_band_std": 0.0012055337535602764, + "matched_frac_signed_per_threshold": { + "0.5": 0.3595106336805556, + "0.6": 0.40082465277777773, + "0.7": 0.444580078125, + "0.8": 0.47905815972222227, + "0.9": 0.4430338541666667 + }, + "matched_frac_signed_band": 0.4418741861979167, + "matched_frac_signed_band_std": 0.003188141718741717, + "babel_mu1_abs_mean": 0.6668546994527181, + "babel_mu1_abs_std": 0.0033598266343399697, + "n_dup_abs_per_threshold": { + "0.5": 8575.0, + "0.6": 7539.666666666667, + "0.7": 6079.333333333333, + "0.8": 3855.3333333333335, + "0.9": 1184.0 + }, + "n_dup_abs_band": 4664.583333333333, + "n_dup_abs_band_std": 35.63033585771166, + "n_dup_clusters_abs_per_threshold": { + "0.5": 2352.6666666666665, + "0.6": 2276.0, + "0.7": 2020.0, + "0.8": 1412.3333333333333, + "0.9": 498.6666666666667 + }, + "n_dup_clusters_abs_band": 1551.75, + "n_dup_clusters_abs_band_std": 10.5, + "largest_dup_abs_per_threshold": { + "0.5": 315.0, + "0.6": 251.33333333333334, + "0.7": 116.33333333333333, + "0.8": 38.333333333333336, + "0.9": 10.0 + }, + "largest_dup_abs_band": 104.0, + "largest_dup_abs_band_std": 8.54400374531753, + "dcs_abs_mean": 0.7811335126558939, + "dcs_abs_std": 0.0023305099951428116, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.3581081847441965, + "0.6": 0.4177895629480999, + "0.7": 0.49546633165462284, + "0.8": 0.6052459945519028, + "0.9": 0.7299357180583886 + }, + "dedup_dcs_abs_band": 0.5621094018032535, + "dedup_dcs_abs_band_std": 0.0009540964043521437, + "matched_cos_abs_per_threshold": { + "0.5": 0.7254678209622701, + "0.6": 0.7308622002601624, + "0.7": 0.7398896018664042, + "0.8": 0.7554369370142618, + "0.9": 0.7730513215065002 + }, + "matched_cos_abs_band": 0.7498100151618321, + "matched_cos_abs_band_std": 0.0020075506852886187, + "unique_frac_abs_per_threshold": { + "0.5": 0.4936252170138889, + "0.6": 0.5716417100694445, + "0.7": 0.669650607638889, + "0.8": 0.8011881510416666, + "0.9": 0.9442274305555557 + }, + "unique_frac_abs_band": 0.746676974826389, + "unique_frac_abs_band_std": 0.0010942126864736965, + "matched_frac_abs_per_threshold": { + "0.5": 0.3592936197916667, + "0.6": 0.4004991319444444, + "0.7": 0.4444173177083333, + "0.8": 0.4796006944444444, + "0.9": 0.44316948784722227 + }, + "matched_frac_abs_band": 0.4419216579861111, + "matched_frac_abs_band_std": 0.003226171902808814, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/seed0/config.json b/vision/dinov2-b_layer10/mp/x16/l0_32/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..582c8dce741510337aae6bc33e8db9b021f25bc1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/seed0/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..8362eb3e73c18d7ca879ba894c4f1d98d685af85 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1298, + "r2_score": 0.9648, + "l0_per_token_0": 0.0029, + "l0_per_token_0.001": 0.0029, + "l0_per_token_0.01": 0.0029, + "l0_per_token_0.05": 0.0029, + "l0_per_token_0.1": 0.0029, + "l0_per_token_0.5": 0.0029, + "l0_per_token_1.0": 0.0026, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 96.7825 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/seed1/config.json b/vision/dinov2-b_layer10/mp/x16/l0_32/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8465153a9e9a1f2031f87935b38abdc81b8f3232 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/seed1/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..53e35fd12396d401b0aad7c91327037b5e6a76a8 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1258, + "r2_score": 0.9659, + "l0_per_token_0": 0.0031, + "l0_per_token_0.001": 0.0031, + "l0_per_token_0.01": 0.0031, + "l0_per_token_0.05": 0.0031, + "l0_per_token_0.1": 0.0031, + "l0_per_token_0.5": 0.0031, + "l0_per_token_1.0": 0.0027, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 100.9962 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/seed2/config.json b/vision/dinov2-b_layer10/mp/x16/l0_32/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..859d3bae315c9ef5dc1a4ece37711d5e66c114f1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/seed2/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..e904d1368bf2bf3804acc8a9aa983626ad162ba9 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1287, + "r2_score": 0.9651, + "l0_per_token_0": 0.003, + "l0_per_token_0.001": 0.003, + "l0_per_token_0.01": 0.003, + "l0_per_token_0.05": 0.003, + "l0_per_token_0.1": 0.003, + "l0_per_token_0.5": 0.003, + "l0_per_token_1.0": 0.0026, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 98.9773 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/code_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x16/l0_64/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..854b8ae336881077d46322bd7310278687f599ff --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 303.350046692607, + "recon_r2": 0.9945499165194519, + "l1_per_token": 235.00570939926905 + }, + "seed1": { + "l0": 308.26556420233464, + "recon_r2": 0.9947738428369166, + "l1_per_token": 237.94410117398962 + }, + "seed2": { + "l0": 303.32809208819714, + "recon_r2": 0.994642080225655, + "l1_per_token": 236.53932976929894 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.3081740635698948, + "mcc_abs": 0.308183846174609, + "act_iou": 0.06158082759211115, + "act_iou_n_pairs": 12288, + "input_r2": 0.9944915843011223, + "matched_mcc_signed": { + "0.6": 0.486311554204309, + "0.7": 0.47742297952236057, + "0.8": 0.4724732855685592, + "0.9": 0.499113829069671 + }, + "matched_mcc_abs": { + "0.6": 0.48502962515759485, + "0.7": 0.47719360709953557, + "0.8": 0.4724564351388078, + "0.9": 0.49935382096494135 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.30419426695625673, + "mcc_abs": 0.30420468250052374, + "act_iou": 0.06069953939073782, + "act_iou_n_pairs": 12288, + "input_r2": 0.994404892243357, + "matched_mcc_signed": { + "0.6": 0.48488945895927105, + "0.7": 0.47440591114644926, + "0.8": 0.4711649853133228, + "0.9": 0.49944654239146674 + }, + "matched_mcc_abs": { + "0.6": 0.4840749925413885, + "0.7": 0.4744411480351126, + "0.8": 0.46963779513166715, + "0.9": 0.4997708880092602 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.3070745853473141, + "mcc_abs": 0.3070853782622768, + "act_iou": 0.06065747285174358, + "act_iou_n_pairs": 12288, + "input_r2": 0.994486634330592, + "matched_mcc_signed": { + "0.6": 0.4903236837017785, + "0.7": 0.48064124130373964, + "0.8": 0.4743521280493735, + "0.9": 0.4958878012745798 + }, + "matched_mcc_abs": { + "0.6": 0.48972885496687907, + "0.7": 0.4805835291198603, + "0.8": 0.4730622072347284, + "0.9": 0.49558014937754513 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 304.9812343277129, + "l0_std": 2.8443342885749576, + "recon_r2_mean": 0.9946552798606746, + "recon_r2_std": 0.00011254519841739937, + "l1_per_token_mean": 236.49638011418588, + "l1_per_token_std": 1.4696666493093344, + "input_r2_mean": 0.9944610369583572, + "input_r2_std": 4.868569932631574e-05, + "mcc_signed_mean": 0.30648097195782187, + "mcc_signed_std": 0.0020552318393844463, + "mcc_abs_mean": 0.30649130231246985, + "mcc_abs_std": 0.0020550256839986587, + "matched_mcc_signed_per_threshold": { + "0.6": 0.4871748989551195, + "0.7": 0.4774900439908498, + "0.8": 0.4726634663104185, + "0.9": 0.49814939091190585 + }, + "matched_mcc_signed_band": 0.48386945004207343, + "matched_mcc_abs_per_threshold": { + "0.6": 0.48627782422195415, + "0.7": 0.47740609475150286, + "0.8": 0.4717188125017344, + "0.9": 0.49823495278391555 + }, + "matched_mcc_abs_band": 0.48340942106477675, + "act_iou_mean": 0.060979279944864184, + "act_iou_std": 0.0005213799740659474, + "act_iou_n_pairs": 12288.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/dict_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x16/l0_64/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..8cd42332d4b9868b8200c9463fc9dca191f3c031 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/dict_metrics_step_final.json @@ -0,0 +1,545 @@ +{ + "schema": 2, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8607, + "0.6": 7559, + "0.7": 6051, + "0.8": 3854, + "0.9": 1228 + }, + "n_dup_clusters_signed": { + "0.5": 2371, + "0.6": 2294, + "0.7": 2002, + "0.8": 1407, + "0.9": 519 + }, + "largest_dup_signed": { + "0.5": 367, + "0.6": 236, + "0.7": 134, + "0.8": 29, + "0.9": 10 + }, + "babel_mu1_signed": 0.6670389771461487, + "n_dup_abs": { + "0.5": 8643, + "0.6": 7607, + "0.7": 6081, + "0.8": 3874, + "0.9": 1228 + }, + "n_dup_clusters_abs": { + "0.5": 2386, + "0.6": 2311, + "0.7": 2013, + "0.8": 1414, + "0.9": 517 + }, + "largest_dup_abs": { + "0.5": 372, + "0.6": 295, + "0.7": 114, + "0.8": 29, + "0.9": 10 + }, + "babel_mu1_abs": 0.6695666313171387 + }, + "seed1": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8532, + "0.6": 7493, + "0.7": 6077, + "0.8": 3854, + "0.9": 1170 + }, + "n_dup_clusters_signed": { + "0.5": 2322, + "0.6": 2234, + "0.7": 2022, + "0.8": 1405, + "0.9": 494 + }, + "largest_dup_signed": { + "0.5": 439, + "0.6": 325, + "0.7": 145, + "0.8": 50, + "0.9": 10 + }, + "babel_mu1_signed": 0.6654933094978333, + "n_dup_abs": { + "0.5": 8570, + "0.6": 7525, + "0.7": 6107, + "0.8": 3874, + "0.9": 1172 + }, + "n_dup_clusters_abs": { + "0.5": 2326, + "0.6": 2245, + "0.7": 2036, + "0.8": 1414, + "0.9": 494 + }, + "largest_dup_abs": { + "0.5": 235, + "0.6": 173, + "0.7": 147, + "0.8": 50, + "0.9": 10 + }, + "babel_mu1_abs": 0.667901337146759 + }, + "seed2": { + "d_sae": 12288, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 8456, + "0.6": 7442, + "0.7": 6006, + "0.8": 3793, + "0.9": 1148 + }, + "n_dup_clusters_signed": { + "0.5": 2327, + "0.6": 2258, + "0.7": 1995, + "0.8": 1400, + "0.9": 485 + }, + "largest_dup_signed": { + "0.5": 355, + "0.6": 144, + "0.7": 77, + "0.8": 36, + "0.9": 10 + }, + "babel_mu1_signed": 0.6603737473487854, + "n_dup_abs": { + "0.5": 8512, + "0.6": 7487, + "0.7": 6050, + "0.8": 3818, + "0.9": 1152 + }, + "n_dup_clusters_abs": { + "0.5": 2346, + "0.6": 2272, + "0.7": 2011, + "0.8": 1409, + "0.9": 485 + }, + "largest_dup_abs": { + "0.5": 338, + "0.6": 286, + "0.7": 88, + "0.8": 36, + "0.9": 10 + }, + "babel_mu1_abs": 0.6630961298942566 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.7817156910896301, + "dcs_abs": 0.7834139466285706, + "frac_sign_flipped": 0.039388020833333336, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.7236672043800354, + "0.6": 0.7293717265129089, + "0.7": 0.7402892112731934, + "0.8": 0.7550824880599976, + "0.9": 0.7738751173019409 + }, + "unique_frac_signed": { + "0.5": 0.4935709635416667, + "0.6": 0.57177734375, + "0.7": 0.6702473958333334, + "0.8": 0.80078125, + "0.9": 0.9436442057291666 + }, + "dedup_dcs_signed": { + "0.5": 0.3571811193493583, + "0.6": 0.4170382283919025, + "0.7": 0.4961769160193702, + "0.8": 0.6046558986417949, + "0.9": 0.7302627703999557 + }, + "matched_frac_signed": { + "0.5": 0.3589680989583333, + "0.6": 0.4026692708333333, + "0.7": 0.4476725260416667, + "0.8": 0.482421875, + "0.9": 0.443359375 + }, + "matched_cos_abs": { + "0.5": 0.7263489365577698, + "0.6": 0.732623279094696, + "0.7": 0.7427051663398743, + "0.8": 0.7570679783821106, + "0.9": 0.7756255269050598 + }, + "unique_frac_abs": { + "0.5": 0.4913330078125, + "0.6": 0.5696614583333334, + "0.7": 0.6688232421875, + "0.8": 0.7998046875, + "0.9": 0.9434814453125 + }, + "dedup_dcs_abs": { + "0.5": 0.35687920772033976, + "0.6": 0.41734724557803327, + "0.7": 0.4967384773408412, + "0.8": 0.6055065178661607, + "0.9": 0.7317882931456552 + }, + "matched_frac_abs": { + "0.5": 0.3590494791666667, + "0.6": 0.4029947916666667, + "0.7": 0.447998046875, + "0.8": 0.4825032552083333, + "0.9": 0.443359375 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.7770745754241943, + "dcs_abs": 0.7787559628486633, + "frac_sign_flipped": 0.043701171875, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.7226144671440125, + "0.6": 0.7275781631469727, + "0.7": 0.7358648180961609, + "0.8": 0.7517416477203369, + "0.9": 0.7691615223884583 + }, + "unique_frac_signed": { + "0.5": 0.4968668619791667, + "0.6": 0.5748291015625, + "0.7": 0.6720377604166666, + "0.8": 0.8030598958333334, + "0.9": 0.9441731770833334 + }, + "dedup_dcs_signed": { + "0.5": 0.3590431827105931, + "0.6": 0.41823310183826834, + "0.7": 0.49452894432276173, + "0.8": 0.6036935693118721, + "0.9": 0.726221678283764 + }, + "matched_frac_signed": { + "0.5": 0.3600260416666667, + "0.6": 0.4007161458333333, + "0.7": 0.4436848958333333, + "0.8": 0.4747721354166667, + "0.9": 0.4336751302083333 + }, + "matched_cos_abs": { + "0.5": 0.7251178026199341, + "0.6": 0.7300339341163635, + "0.7": 0.737865686416626, + "0.8": 0.7535355687141418, + "0.9": 0.7708365321159363 + }, + "unique_frac_abs": { + "0.5": 0.4945068359375, + "0.6": 0.5723063151041666, + "0.7": 0.6701253255208334, + "0.8": 0.8018798828125, + "0.9": 0.9439290364583334 + }, + "dedup_dcs_abs": { + "0.5": 0.35857571025553625, + "0.6": 0.417803030735134, + "0.7": 0.4944624833005946, + "0.8": 0.6042450135355466, + "0.9": 0.7276149850270789 + }, + "matched_frac_abs": { + "0.5": 0.3597005208333333, + "0.6": 0.4002278645833333, + "0.7": 0.443115234375, + "0.8": 0.475830078125, + "0.9": 0.4337565104166667 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.7794721722602844, + "dcs_abs": 0.781230628490448, + "frac_sign_flipped": 0.042073567708333336, + "alive_source": "stats", + "n_alive_a": 12288, + "n_alive_b": 12288, + "matched_cos_signed": { + "0.5": 0.722134530544281, + "0.6": 0.7282166481018066, + "0.7": 0.7373058795928955, + "0.8": 0.7540035247802734, + "0.9": 0.7709962725639343 + }, + "unique_frac_signed": { + "0.5": 0.4979248046875, + "0.6": 0.5750732421875, + "0.7": 0.6717936197916666, + "0.8": 0.802978515625, + "0.9": 0.9455159505208334 + }, + "dedup_dcs_signed": { + "0.5": 0.3595686950793606, + "0.6": 0.4187779088388197, + "0.7": 0.49531738574538997, + "0.8": 0.6054486311040819, + "0.9": 0.7289892735013079 + }, + "matched_frac_signed": { + "0.5": 0.3595377604166667, + "0.6": 0.3990885416666667, + "0.7": 0.4423828125, + "0.8": 0.47998046875, + "0.9": 0.4520670572916667 + }, + "matched_cos_abs": { + "0.5": 0.7249367237091064, + "0.6": 0.7299293875694275, + "0.7": 0.7390979528427124, + "0.8": 0.7557072639465332, + "0.9": 0.7726919054985046 + }, + "unique_frac_abs": { + "0.5": 0.4950358072916667, + "0.6": 0.5729573567708334, + "0.7": 0.6700032552083334, + "0.8": 0.8018798828125, + "0.9": 0.9452718098958334 + }, + "dedup_dcs_abs": { + "0.5": 0.3588696362567134, + "0.6": 0.41821841253113234, + "0.7": 0.4951980343224326, + "0.8": 0.605986452254001, + "0.9": 0.7304038760024317 + }, + "matched_frac_abs": { + "0.5": 0.359130859375, + "0.6": 0.3982747395833333, + "0.7": 0.442138671875, + "0.8": 0.48046875, + "0.9": 0.452392578125 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 0, + "n_dense_std": 0.0, + "l0_mean": null, + "l0_std": null, + "recon_r2_mean": null, + "recon_r2_std": null, + "l1_per_token_mean": null, + "l1_per_token_std": null, + "babel_mu1_signed_mean": 0.6643020113309225, + "babel_mu1_signed_std": 0.0034886552645365565, + "n_dup_signed_per_threshold": { + "0.5": 8531.666666666666, + "0.6": 7498.0, + "0.7": 6044.666666666667, + "0.8": 3833.6666666666665, + "0.9": 1182.0 + }, + "n_dup_signed_band": 4639.583333333334, + "n_dup_signed_band_std": 38.65418261111381, + "n_dup_clusters_signed_per_threshold": { + "0.5": 2340.0, + "0.6": 2262.0, + "0.7": 2006.3333333333333, + "0.8": 1404.0, + "0.9": 499.3333333333333 + }, + "n_dup_clusters_signed_band": 1542.9166666666665, + "n_dup_clusters_signed_band_std": 11.10273990208423, + "largest_dup_signed_per_threshold": { + "0.5": 387.0, + "0.6": 235.0, + "0.7": 118.66666666666667, + "0.8": 38.333333333333336, + "0.9": 10.0 + }, + "largest_dup_signed_band": 100.5, + "largest_dup_signed_band_std": 32.90991491936739, + "dcs_signed_mean": 0.779420812924703, + "dcs_signed_std": 0.0023209840566022644, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.35859766571310403, + "0.6": 0.41801641302299686, + "0.7": 0.49534108202917393, + "0.8": 0.6045993663525829, + "0.9": 0.7284912407283425 + }, + "dedup_dcs_signed_band": 0.561612025533274, + "dedup_dcs_signed_band_std": 0.000817928943294489, + "matched_cos_signed_per_threshold": { + "0.5": 0.722805400689443, + "0.6": 0.7283888459205627, + "0.7": 0.7378199696540833, + "0.8": 0.7536092201868693, + "0.9": 0.7713443040847778 + }, + "matched_cos_signed_band": 0.7477905849615732, + "matched_cos_signed_band_std": 0.001789422150281412, + "unique_frac_signed_per_threshold": { + "0.5": 0.49612087673611116, + "0.6": 0.5738932291666666, + "0.7": 0.6713595920138888, + "0.8": 0.8022732204861112, + "0.9": 0.9444444444444445 + }, + "unique_frac_signed_band": 0.7479926215277778, + "unique_frac_signed_band_std": 0.0012055337535602764, + "matched_frac_signed_per_threshold": { + "0.5": 0.3595106336805556, + "0.6": 0.40082465277777773, + "0.7": 0.444580078125, + "0.8": 0.47905815972222227, + "0.9": 0.4430338541666667 + }, + "matched_frac_signed_band": 0.4418741861979167, + "matched_frac_signed_band_std": 0.003188141718741717, + "babel_mu1_abs_mean": 0.6668546994527181, + "babel_mu1_abs_std": 0.0033598266343399697, + "n_dup_abs_per_threshold": { + "0.5": 8575.0, + "0.6": 7539.666666666667, + "0.7": 6079.333333333333, + "0.8": 3855.3333333333335, + "0.9": 1184.0 + }, + "n_dup_abs_band": 4664.583333333333, + "n_dup_abs_band_std": 35.63033585771166, + "n_dup_clusters_abs_per_threshold": { + "0.5": 2352.6666666666665, + "0.6": 2276.0, + "0.7": 2020.0, + "0.8": 1412.3333333333333, + "0.9": 498.6666666666667 + }, + "n_dup_clusters_abs_band": 1551.75, + "n_dup_clusters_abs_band_std": 10.5, + "largest_dup_abs_per_threshold": { + "0.5": 315.0, + "0.6": 251.33333333333334, + "0.7": 116.33333333333333, + "0.8": 38.333333333333336, + "0.9": 10.0 + }, + "largest_dup_abs_band": 104.0, + "largest_dup_abs_band_std": 8.54400374531753, + "dcs_abs_mean": 0.7811335126558939, + "dcs_abs_std": 0.0023305099951428116, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.3581081847441965, + "0.6": 0.4177895629480999, + "0.7": 0.49546633165462284, + "0.8": 0.6052459945519028, + "0.9": 0.7299357180583886 + }, + "dedup_dcs_abs_band": 0.5621094018032535, + "dedup_dcs_abs_band_std": 0.0009540964043521437, + "matched_cos_abs_per_threshold": { + "0.5": 0.7254678209622701, + "0.6": 0.7308622002601624, + "0.7": 0.7398896018664042, + "0.8": 0.7554369370142618, + "0.9": 0.7730513215065002 + }, + "matched_cos_abs_band": 0.7498100151618321, + "matched_cos_abs_band_std": 0.0020075506852886187, + "unique_frac_abs_per_threshold": { + "0.5": 0.4936252170138889, + "0.6": 0.5716417100694445, + "0.7": 0.669650607638889, + "0.8": 0.8011881510416666, + "0.9": 0.9442274305555557 + }, + "unique_frac_abs_band": 0.746676974826389, + "unique_frac_abs_band_std": 0.0010942126864736965, + "matched_frac_abs_per_threshold": { + "0.5": 0.3592936197916667, + "0.6": 0.4004991319444444, + "0.7": 0.4444173177083333, + "0.8": 0.4796006944444444, + "0.9": 0.44316948784722227 + }, + "matched_frac_abs_band": 0.4419216579861111, + "matched_frac_abs_band_std": 0.003226171902808814, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/seed0/config.json b/vision/dinov2-b_layer10/mp/x16/l0_64/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..582c8dce741510337aae6bc33e8db9b021f25bc1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/seed0/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..8362eb3e73c18d7ca879ba894c4f1d98d685af85 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1298, + "r2_score": 0.9648, + "l0_per_token_0": 0.0029, + "l0_per_token_0.001": 0.0029, + "l0_per_token_0.01": 0.0029, + "l0_per_token_0.05": 0.0029, + "l0_per_token_0.1": 0.0029, + "l0_per_token_0.5": 0.0029, + "l0_per_token_1.0": 0.0026, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 96.7825 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/seed1/config.json b/vision/dinov2-b_layer10/mp/x16/l0_64/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8465153a9e9a1f2031f87935b38abdc81b8f3232 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/seed1/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..53e35fd12396d401b0aad7c91327037b5e6a76a8 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1258, + "r2_score": 0.9659, + "l0_per_token_0": 0.0031, + "l0_per_token_0.001": 0.0031, + "l0_per_token_0.01": 0.0031, + "l0_per_token_0.05": 0.0031, + "l0_per_token_0.1": 0.0031, + "l0_per_token_0.5": 0.0031, + "l0_per_token_1.0": 0.0027, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 100.9962 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/seed2/config.json b/vision/dinov2-b_layer10/mp/x16/l0_64/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..859d3bae315c9ef5dc1a4ece37711d5e66c114f1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/seed2/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "k_iters": 10, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 10, + "residual_thr": 0.05, + "max_iters": 1228, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..e904d1368bf2bf3804acc8a9aa983626ad162ba9 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1287, + "r2_score": 0.9651, + "l0_per_token_0": 0.003, + "l0_per_token_0.001": 0.003, + "l0_per_token_0.01": 0.003, + "l0_per_token_0.05": 0.003, + "l0_per_token_0.1": 0.003, + "l0_per_token_0.5": 0.003, + "l0_per_token_1.0": 0.0026, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 98.9773 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_128/code_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x4/l0_128/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..641ddb41a29bb4274dfab5a4c5ae30936d7cd404 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_128/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 94.26257976653696, + "recon_r2": 0.975821934346816, + "l1_per_token": 160.12352198927357 + }, + "seed1": { + "l0": 96.0824747081712, + "recon_r2": 0.9758979026055418, + "l1_per_token": 160.55607186960745 + }, + "seed2": { + "l0": 96.16374837872893, + "recon_r2": 0.9758934109302257, + "l1_per_token": 160.4021775571133 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.686996152798104, + "mcc_abs": 0.6870703592996316, + "act_iou": 0.25547694536790305, + "act_iou_n_pairs": 3072, + "input_r2": 0.9846564578689908, + "matched_mcc_signed": { + "0.6": 0.8233672574692353, + "0.7": 0.8235882716727134, + "0.8": 0.8303783596636519, + "0.9": 0.8476126278762411 + }, + "matched_mcc_abs": { + "0.6": 0.8221728981196199, + "0.7": 0.822716947685233, + "0.8": 0.831733380016577, + "0.9": 0.8478334868152545 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6846840780527694, + "mcc_abs": 0.6847622419526243, + "act_iou": 0.25246337038642785, + "act_iou_n_pairs": 3072, + "input_r2": 0.9844104815350083, + "matched_mcc_signed": { + "0.6": 0.8227527574304134, + "0.7": 0.8178582045636273, + "0.8": 0.8274592190546871, + "0.9": 0.844338840962079 + }, + "matched_mcc_abs": { + "0.6": 0.8202325549355041, + "0.7": 0.8191295136242774, + "0.8": 0.8274661268153322, + "0.9": 0.8450663089562113 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6946268447282802, + "mcc_abs": 0.6946844550087684, + "act_iou": 0.2577866682674802, + "act_iou_n_pairs": 3072, + "input_r2": 0.9846777056136713, + "matched_mcc_signed": { + "0.6": 0.822279493121979, + "0.7": 0.818058897473975, + "0.8": 0.8203782644254983, + "0.9": 0.8269563531697791 + }, + "matched_mcc_abs": { + "0.6": 0.820086935363402, + "0.7": 0.8187139526860329, + "0.8": 0.819773872370219, + "0.9": 0.8279184476094364 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 95.50293428447902, + "l0_std": 1.0749469056730019, + "recon_r2_mean": 0.975871082627528, + "recon_r2_std": 4.2622868386407093e-05, + "l1_per_token_mean": 160.3605904719981, + "l1_per_token_std": 0.2192531960553768, + "input_r2_mean": 0.9845815483392236, + "input_r2_std": 0.00014852863458536, + "mcc_signed_mean": 0.6887690251930513, + "mcc_signed_std": 0.005203072139369793, + "mcc_abs_mean": 0.6888390187536748, + "mcc_abs_std": 0.005192176343419167, + "matched_mcc_signed_per_threshold": { + "0.6": 0.8227998360072094, + "0.7": 0.8198351245701052, + "0.8": 0.8260719477146125, + "0.9": 0.8396359406693664 + }, + "matched_mcc_signed_band": 0.8270857122403233, + "matched_mcc_abs_per_threshold": { + "0.6": 0.8208307961395086, + "0.7": 0.8201868046651811, + "0.8": 0.8263244597340428, + "0.9": 0.8402727477936341 + }, + "matched_mcc_abs_band": 0.8269037020830917, + "act_iou_mean": 0.25524232800727037, + "act_iou_std": 0.002669393013783618, + "act_iou_n_pairs": 3072.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_128/dict_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x4/l0_128/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..3cc4fa639964086adc9fb5a8440b0d7fe5b772b1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_128/dict_metrics_step_final.json @@ -0,0 +1,595 @@ +{ + "schema": 3, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 633, + "0.6": 505, + "0.7": 383, + "0.8": 264, + "0.9": 129 + }, + "n_dup_clusters_signed": { + "0.5": 178, + "0.6": 140, + "0.7": 110, + "0.8": 71, + "0.9": 33 + }, + "n_dup_removed_signed": { + "0.5": 455, + "0.6": 365, + "0.7": 273, + "0.8": 193, + "0.9": 96 + }, + "largest_dup_signed": { + "0.5": 102, + "0.6": 56, + "0.7": 46, + "0.8": 27, + "0.9": 14 + }, + "babel_mu1_signed": 0.3663797080516815, + "n_dup_abs": { + "0.5": 644, + "0.6": 511, + "0.7": 389, + "0.8": 267, + "0.9": 130 + }, + "n_dup_clusters_abs": { + "0.5": 181, + "0.6": 142, + "0.7": 112, + "0.8": 72, + "0.9": 34 + }, + "n_dup_removed_abs": { + "0.5": 463, + "0.6": 369, + "0.7": 277, + "0.8": 195, + "0.9": 96 + }, + "largest_dup_abs": { + "0.5": 106, + "0.6": 55, + "0.7": 51, + "0.8": 27, + "0.9": 11 + }, + "babel_mu1_abs": 0.375279039144516, + "dup_excludes_dead": true + }, + "seed1": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 658, + "0.6": 541, + "0.7": 413, + "0.8": 303, + "0.9": 198 + }, + "n_dup_clusters_signed": { + "0.5": 174, + "0.6": 140, + "0.7": 100, + "0.8": 65, + "0.9": 60 + }, + "n_dup_removed_signed": { + "0.5": 484, + "0.6": 401, + "0.7": 313, + "0.8": 238, + "0.9": 138 + }, + "largest_dup_signed": { + "0.5": 89, + "0.6": 48, + "0.7": 28, + "0.8": 22, + "0.9": 10 + }, + "babel_mu1_signed": 0.3705531060695648, + "n_dup_abs": { + "0.5": 671, + "0.6": 548, + "0.7": 414, + "0.8": 308, + "0.9": 200 + }, + "n_dup_clusters_abs": { + "0.5": 179, + "0.6": 140, + "0.7": 97, + "0.8": 67, + "0.9": 61 + }, + "n_dup_removed_abs": { + "0.5": 492, + "0.6": 408, + "0.7": 317, + "0.8": 241, + "0.9": 139 + }, + "largest_dup_abs": { + "0.5": 73, + "0.6": 70, + "0.7": 28, + "0.8": 22, + "0.9": 10 + }, + "babel_mu1_abs": 0.3797185719013214, + "dup_excludes_dead": true + }, + "seed2": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 639, + "0.6": 540, + "0.7": 423, + "0.8": 318, + "0.9": 208 + }, + "n_dup_clusters_signed": { + "0.5": 156, + "0.6": 128, + "0.7": 99, + "0.8": 74, + "0.9": 63 + }, + "n_dup_removed_signed": { + "0.5": 483, + "0.6": 412, + "0.7": 324, + "0.8": 244, + "0.9": 145 + }, + "largest_dup_signed": { + "0.5": 111, + "0.6": 57, + "0.7": 56, + "0.8": 45, + "0.9": 14 + }, + "babel_mu1_signed": 0.3651764392852783, + "n_dup_abs": { + "0.5": 642, + "0.6": 548, + "0.7": 429, + "0.8": 324, + "0.9": 212 + }, + "n_dup_clusters_abs": { + "0.5": 158, + "0.6": 133, + "0.7": 103, + "0.8": 78, + "0.9": 65 + }, + "n_dup_removed_abs": { + "0.5": 484, + "0.6": 415, + "0.7": 326, + "0.8": 246, + "0.9": 147 + }, + "largest_dup_abs": { + "0.5": 62, + "0.6": 57, + "0.7": 56, + "0.8": 45, + "0.9": 14 + }, + "babel_mu1_abs": 0.37334656715393066, + "dup_excludes_dead": true + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.863697350025177, + "dcs_abs": 0.8660604953765869, + "frac_sign_flipped": 0.033528645833333336, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8904803991317749, + "0.6": 0.8863230347633362, + "0.7": 0.8811215758323669, + "0.8": 0.8779223561286926, + "0.9": 0.8697978258132935 + }, + "unique_frac_signed": { + "0.5": 0.84716796875, + "0.6": 0.8753255208333334, + "0.7": 0.9046223958333334, + "0.8": 0.9298502604166666, + "0.9": 0.9619140625 + }, + "dedup_dcs_signed": { + "0.5": 0.754386470944155, + "0.6": 0.7758211720307979, + "0.7": 0.7970823109499179, + "0.8": 0.8163363314718783, + "0.9": 0.8366707601817325 + }, + "matched_frac_signed": { + "0.5": 0.7802734375, + "0.6": 0.7838541666666666, + "0.7": 0.7750651041666666, + "0.8": 0.7529296875, + "0.9": 0.6888020833333334 + }, + "matched_cos_abs": { + "0.5": 0.892687201499939, + "0.6": 0.8885923027992249, + "0.7": 0.8832919001579285, + "0.8": 0.8800186514854431, + "0.9": 0.8724757432937622 + }, + "unique_frac_abs": { + "0.5": 0.8445638020833334, + "0.6": 0.87353515625, + "0.7": 0.9033203125, + "0.8": 0.9290364583333334, + "0.9": 0.9617513020833334 + }, + "dedup_dcs_abs": { + "0.5": 0.7539312969699192, + "0.6": 0.7762166160682682, + "0.7": 0.7978955152793787, + "0.8": 0.8175694112433121, + "0.9": 0.8391046821488999 + }, + "matched_frac_abs": { + "0.5": 0.7802734375, + "0.6": 0.7848307291666666, + "0.7": 0.7760416666666666, + "0.8": 0.751953125, + "0.9": 0.6884765625 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.8620738983154297, + "dcs_abs": 0.8638548851013184, + "frac_sign_flipped": 0.028645833333333332, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8867824077606201, + "0.6": 0.8854392170906067, + "0.7": 0.8801518082618713, + "0.8": 0.8756371140480042, + "0.9": 0.8707349896430969 + }, + "unique_frac_signed": { + "0.5": 0.8473307291666666, + "0.6": 0.87353515625, + "0.7": 0.90283203125, + "0.8": 0.9288736979166666, + "0.9": 0.9607747395833334 + }, + "dedup_dcs_signed": { + "0.5": 0.7513979841799785, + "0.6": 0.7734622848511208, + "0.7": 0.7946292448614258, + "0.8": 0.8133562841588475, + "0.9": 0.8365801829204429 + }, + "matched_frac_signed": { + "0.5": 0.7734375, + "0.6": 0.7809244791666666, + "0.7": 0.7770182291666666, + "0.8": 0.7513020833333334, + "0.9": 0.6901041666666666 + }, + "matched_cos_abs": { + "0.5": 0.8874406218528748, + "0.6": 0.8861882090568542, + "0.7": 0.8809804916381836, + "0.8": 0.8767107129096985, + "0.9": 0.8725345730781555 + }, + "unique_frac_abs": { + "0.5": 0.8458658854166666, + "0.6": 0.8723958333333334, + "0.7": 0.90185546875, + "0.8": 0.92822265625, + "0.9": 0.96044921875 + }, + "dedup_dcs_abs": { + "0.5": 0.7506557473582991, + "0.6": 0.7731069011303286, + "0.7": 0.7945170742459595, + "0.8": 0.8137827466998715, + "0.9": 0.8380251490452792 + }, + "matched_frac_abs": { + "0.5": 0.7737630208333334, + "0.6": 0.7822265625, + "0.7": 0.7760416666666666, + "0.8": 0.7506510416666666, + "0.9": 0.6888020833333334 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.8889439702033997, + "dcs_abs": 0.8901732563972473, + "frac_sign_flipped": 0.0185546875, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8958790898323059, + "0.6": 0.893430769443512, + "0.7": 0.889135479927063, + "0.8": 0.8864941000938416, + "0.9": 0.8865686058998108 + }, + "unique_frac_signed": { + "0.5": 0.8426106770833334, + "0.6": 0.86767578125, + "0.7": 0.8963216145833334, + "0.8": 0.9215494791666666, + "0.9": 0.9539388020833334 + }, + "dedup_dcs_signed": { + "0.5": 0.7548772864683997, + "0.6": 0.7752082408696879, + "0.7": 0.7969513489515521, + "0.8": 0.8169481762258025, + "0.9": 0.8457321938767564 + }, + "matched_frac_signed": { + "0.5": 0.7760416666666666, + "0.6": 0.7858072916666666, + "0.7": 0.78515625, + "0.8": 0.7705078125, + "0.9": 0.7272135416666666 + }, + "matched_cos_abs": { + "0.5": 0.8988944888114929, + "0.6": 0.8942874073982239, + "0.7": 0.8906096816062927, + "0.8": 0.8878304362297058, + "0.9": 0.8880638480186462 + }, + "unique_frac_abs": { + "0.5": 0.8411458333333334, + "0.6": 0.8660481770833334, + "0.7": 0.8953450520833334, + "0.8": 0.9207356770833334, + "0.9": 0.9534505208333334 + }, + "dedup_dcs_abs": { + "0.5": 0.7561013538700839, + "0.6": 0.7744959789658121, + "0.7": 0.7974029717637071, + "0.8": 0.8174571578371495, + "0.9": 0.8467249384266324 + }, + "matched_frac_abs": { + "0.5": 0.77734375, + "0.6": 0.787109375, + "0.7": 0.7854817708333334, + "0.8": 0.7708333333333334, + "0.9": 0.7259114583333334 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 0, + "n_dense_std": 0.0, + "l0_mean": null, + "l0_std": null, + "recon_r2_mean": null, + "recon_r2_std": null, + "l1_per_token_mean": null, + "l1_per_token_std": null, + "babel_mu1_signed_mean": 0.36736975113550824, + "babel_mu1_signed_std": 0.0028217504156440275, + "n_dup_signed_per_threshold": { + "0.5": 643.3333333333334, + "0.6": 528.6666666666666, + "0.7": 406.3333333333333, + "0.8": 295.0, + "0.9": 178.33333333333334 + }, + "n_dup_signed_band": 352.0833333333333, + "n_dup_signed_band_std": 27.894145144336893, + "n_dup_clusters_signed_per_threshold": { + "0.5": 169.33333333333334, + "0.6": 136.0, + "0.7": 103.0, + "0.8": 70.0, + "0.9": 52.0 + }, + "n_dup_clusters_signed_band": 90.25, + "n_dup_clusters_signed_band_std": 1.5206906325745548, + "largest_dup_signed_per_threshold": { + "0.5": 100.66666666666667, + "0.6": 53.666666666666664, + "0.7": 43.333333333333336, + "0.8": 31.333333333333332, + "0.9": 12.666666666666666 + }, + "largest_dup_signed_band": 35.25, + "largest_dup_signed_band_std": 8.011710179481033, + "dcs_signed_mean": 0.8715717395146688, + "dcs_signed_std": 0.015066675087434934, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.7535539138641777, + "0.6": 0.7748305659172022, + "0.7": 0.7962209682542986, + "0.8": 0.8155469306188428, + "0.9": 0.839661045659644 + }, + "dedup_dcs_signed_band": 0.8065648776124968, + "dedup_dcs_signed_band_std": 0.002102852872293367, + "matched_cos_signed_per_threshold": { + "0.5": 0.8910472989082336, + "0.6": 0.8883976737658182, + "0.7": 0.8834696213404337, + "0.8": 0.8800178567568461, + "0.9": 0.8757004737854004 + }, + "matched_cos_signed_band": 0.8818964064121246, + "matched_cos_signed_band_std": 0.006084734579468042, + "unique_frac_signed_per_threshold": { + "0.5": 0.845703125, + "0.6": 0.8721788194444445, + "0.7": 0.9012586805555557, + "0.8": 0.9267578125, + "0.9": 0.9588758680555557 + }, + "unique_frac_signed_band": 0.914767795138889, + "unique_frac_signed_band_std": 0.004299758816805034, + "matched_frac_signed_per_threshold": { + "0.5": 0.7765842013888888, + "0.6": 0.7835286458333334, + "0.7": 0.779079861111111, + "0.8": 0.7582465277777778, + "0.9": 0.7020399305555555 + }, + "matched_frac_signed_band": 0.7557237413194444, + "matched_frac_signed_band_std": 0.009915146747727952, + "babel_mu1_abs_mean": 0.37611472606658936, + "babel_mu1_abs_std": 0.003267168590536138, + "n_dup_abs_per_threshold": { + "0.5": 652.3333333333334, + "0.6": 535.6666666666666, + "0.7": 410.6666666666667, + "0.8": 299.6666666666667, + "0.9": 180.66666666666666 + }, + "n_dup_abs_band": 356.6666666666667, + "n_dup_abs_band_std": 28.583576286625387, + "n_dup_clusters_abs_per_threshold": { + "0.5": 172.66666666666666, + "0.6": 138.33333333333334, + "0.7": 104.0, + "0.8": 72.33333333333333, + "0.9": 53.333333333333336 + }, + "n_dup_clusters_abs_band": 92.0, + "n_dup_clusters_abs_band_std": 2.462214450449026, + "largest_dup_abs_per_threshold": { + "0.5": 80.33333333333333, + "0.6": 60.666666666666664, + "0.7": 45.0, + "0.8": 31.333333333333332, + "0.9": 11.666666666666666 + }, + "largest_dup_abs_band": 37.166666666666664, + "largest_dup_abs_band_std": 5.346338310781814, + "dcs_abs_mean": 0.8733628789583842, + "dcs_abs_std": 0.014599923677861167, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.7535627993994342, + "0.6": 0.7746064987214697, + "0.7": 0.7966051870963485, + "0.8": 0.8162697719267777, + "0.9": 0.8412849232069372 + }, + "dedup_dcs_abs_band": 0.8071915952378833, + "dedup_dcs_abs_band_std": 0.002126596326879785, + "matched_cos_abs_per_threshold": { + "0.5": 0.8930074373881022, + "0.6": 0.889689306418101, + "0.7": 0.8849606911341349, + "0.8": 0.8815199335416158, + "0.9": 0.877691388130188 + }, + "matched_cos_abs_band": 0.8834653298060099, + "matched_cos_abs_band_std": 0.005914915546760715, + "unique_frac_abs_per_threshold": { + "0.5": 0.8438585069444445, + "0.6": 0.8706597222222223, + "0.7": 0.9001736111111112, + "0.8": 0.925998263888889, + "0.9": 0.9585503472222223 + }, + "unique_frac_abs_band": 0.9138454861111112, + "unique_frac_abs_band_std": 0.00432777723906263, + "matched_frac_abs_per_threshold": { + "0.5": 0.7771267361111112, + "0.6": 0.7847222222222222, + "0.7": 0.7791883680555555, + "0.8": 0.7578125, + "0.9": 0.7010633680555557 + }, + "matched_frac_abs_band": 0.7556966145833334, + "matched_frac_abs_band_std": 0.01008819208703062, + "alive_source": "stats", + "frac_denominator": "d_sae", + "d_sae": 3072 + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_32/code_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x4/l0_32/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..641ddb41a29bb4274dfab5a4c5ae30936d7cd404 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_32/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 94.26257976653696, + "recon_r2": 0.975821934346816, + "l1_per_token": 160.12352198927357 + }, + "seed1": { + "l0": 96.0824747081712, + "recon_r2": 0.9758979026055418, + "l1_per_token": 160.55607186960745 + }, + "seed2": { + "l0": 96.16374837872893, + "recon_r2": 0.9758934109302257, + "l1_per_token": 160.4021775571133 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.686996152798104, + "mcc_abs": 0.6870703592996316, + "act_iou": 0.25547694536790305, + "act_iou_n_pairs": 3072, + "input_r2": 0.9846564578689908, + "matched_mcc_signed": { + "0.6": 0.8233672574692353, + "0.7": 0.8235882716727134, + "0.8": 0.8303783596636519, + "0.9": 0.8476126278762411 + }, + "matched_mcc_abs": { + "0.6": 0.8221728981196199, + "0.7": 0.822716947685233, + "0.8": 0.831733380016577, + "0.9": 0.8478334868152545 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6846840780527694, + "mcc_abs": 0.6847622419526243, + "act_iou": 0.25246337038642785, + "act_iou_n_pairs": 3072, + "input_r2": 0.9844104815350083, + "matched_mcc_signed": { + "0.6": 0.8227527574304134, + "0.7": 0.8178582045636273, + "0.8": 0.8274592190546871, + "0.9": 0.844338840962079 + }, + "matched_mcc_abs": { + "0.6": 0.8202325549355041, + "0.7": 0.8191295136242774, + "0.8": 0.8274661268153322, + "0.9": 0.8450663089562113 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6946268447282802, + "mcc_abs": 0.6946844550087684, + "act_iou": 0.2577866682674802, + "act_iou_n_pairs": 3072, + "input_r2": 0.9846777056136713, + "matched_mcc_signed": { + "0.6": 0.822279493121979, + "0.7": 0.818058897473975, + "0.8": 0.8203782644254983, + "0.9": 0.8269563531697791 + }, + "matched_mcc_abs": { + "0.6": 0.820086935363402, + "0.7": 0.8187139526860329, + "0.8": 0.819773872370219, + "0.9": 0.8279184476094364 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 95.50293428447902, + "l0_std": 1.0749469056730019, + "recon_r2_mean": 0.975871082627528, + "recon_r2_std": 4.2622868386407093e-05, + "l1_per_token_mean": 160.3605904719981, + "l1_per_token_std": 0.2192531960553768, + "input_r2_mean": 0.9845815483392236, + "input_r2_std": 0.00014852863458536, + "mcc_signed_mean": 0.6887690251930513, + "mcc_signed_std": 0.005203072139369793, + "mcc_abs_mean": 0.6888390187536748, + "mcc_abs_std": 0.005192176343419167, + "matched_mcc_signed_per_threshold": { + "0.6": 0.8227998360072094, + "0.7": 0.8198351245701052, + "0.8": 0.8260719477146125, + "0.9": 0.8396359406693664 + }, + "matched_mcc_signed_band": 0.8270857122403233, + "matched_mcc_abs_per_threshold": { + "0.6": 0.8208307961395086, + "0.7": 0.8201868046651811, + "0.8": 0.8263244597340428, + "0.9": 0.8402727477936341 + }, + "matched_mcc_abs_band": 0.8269037020830917, + "act_iou_mean": 0.25524232800727037, + "act_iou_std": 0.002669393013783618, + "act_iou_n_pairs": 3072.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_32/dict_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x4/l0_32/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..3cc4fa639964086adc9fb5a8440b0d7fe5b772b1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_32/dict_metrics_step_final.json @@ -0,0 +1,595 @@ +{ + "schema": 3, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 633, + "0.6": 505, + "0.7": 383, + "0.8": 264, + "0.9": 129 + }, + "n_dup_clusters_signed": { + "0.5": 178, + "0.6": 140, + "0.7": 110, + "0.8": 71, + "0.9": 33 + }, + "n_dup_removed_signed": { + "0.5": 455, + "0.6": 365, + "0.7": 273, + "0.8": 193, + "0.9": 96 + }, + "largest_dup_signed": { + "0.5": 102, + "0.6": 56, + "0.7": 46, + "0.8": 27, + "0.9": 14 + }, + "babel_mu1_signed": 0.3663797080516815, + "n_dup_abs": { + "0.5": 644, + "0.6": 511, + "0.7": 389, + "0.8": 267, + "0.9": 130 + }, + "n_dup_clusters_abs": { + "0.5": 181, + "0.6": 142, + "0.7": 112, + "0.8": 72, + "0.9": 34 + }, + "n_dup_removed_abs": { + "0.5": 463, + "0.6": 369, + "0.7": 277, + "0.8": 195, + "0.9": 96 + }, + "largest_dup_abs": { + "0.5": 106, + "0.6": 55, + "0.7": 51, + "0.8": 27, + "0.9": 11 + }, + "babel_mu1_abs": 0.375279039144516, + "dup_excludes_dead": true + }, + "seed1": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 658, + "0.6": 541, + "0.7": 413, + "0.8": 303, + "0.9": 198 + }, + "n_dup_clusters_signed": { + "0.5": 174, + "0.6": 140, + "0.7": 100, + "0.8": 65, + "0.9": 60 + }, + "n_dup_removed_signed": { + "0.5": 484, + "0.6": 401, + "0.7": 313, + "0.8": 238, + "0.9": 138 + }, + "largest_dup_signed": { + "0.5": 89, + "0.6": 48, + "0.7": 28, + "0.8": 22, + "0.9": 10 + }, + "babel_mu1_signed": 0.3705531060695648, + "n_dup_abs": { + "0.5": 671, + "0.6": 548, + "0.7": 414, + "0.8": 308, + "0.9": 200 + }, + "n_dup_clusters_abs": { + "0.5": 179, + "0.6": 140, + "0.7": 97, + "0.8": 67, + "0.9": 61 + }, + "n_dup_removed_abs": { + "0.5": 492, + "0.6": 408, + "0.7": 317, + "0.8": 241, + "0.9": 139 + }, + "largest_dup_abs": { + "0.5": 73, + "0.6": 70, + "0.7": 28, + "0.8": 22, + "0.9": 10 + }, + "babel_mu1_abs": 0.3797185719013214, + "dup_excludes_dead": true + }, + "seed2": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 639, + "0.6": 540, + "0.7": 423, + "0.8": 318, + "0.9": 208 + }, + "n_dup_clusters_signed": { + "0.5": 156, + "0.6": 128, + "0.7": 99, + "0.8": 74, + "0.9": 63 + }, + "n_dup_removed_signed": { + "0.5": 483, + "0.6": 412, + "0.7": 324, + "0.8": 244, + "0.9": 145 + }, + "largest_dup_signed": { + "0.5": 111, + "0.6": 57, + "0.7": 56, + "0.8": 45, + "0.9": 14 + }, + "babel_mu1_signed": 0.3651764392852783, + "n_dup_abs": { + "0.5": 642, + "0.6": 548, + "0.7": 429, + "0.8": 324, + "0.9": 212 + }, + "n_dup_clusters_abs": { + "0.5": 158, + "0.6": 133, + "0.7": 103, + "0.8": 78, + "0.9": 65 + }, + "n_dup_removed_abs": { + "0.5": 484, + "0.6": 415, + "0.7": 326, + "0.8": 246, + "0.9": 147 + }, + "largest_dup_abs": { + "0.5": 62, + "0.6": 57, + "0.7": 56, + "0.8": 45, + "0.9": 14 + }, + "babel_mu1_abs": 0.37334656715393066, + "dup_excludes_dead": true + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.863697350025177, + "dcs_abs": 0.8660604953765869, + "frac_sign_flipped": 0.033528645833333336, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8904803991317749, + "0.6": 0.8863230347633362, + "0.7": 0.8811215758323669, + "0.8": 0.8779223561286926, + "0.9": 0.8697978258132935 + }, + "unique_frac_signed": { + "0.5": 0.84716796875, + "0.6": 0.8753255208333334, + "0.7": 0.9046223958333334, + "0.8": 0.9298502604166666, + "0.9": 0.9619140625 + }, + "dedup_dcs_signed": { + "0.5": 0.754386470944155, + "0.6": 0.7758211720307979, + "0.7": 0.7970823109499179, + "0.8": 0.8163363314718783, + "0.9": 0.8366707601817325 + }, + "matched_frac_signed": { + "0.5": 0.7802734375, + "0.6": 0.7838541666666666, + "0.7": 0.7750651041666666, + "0.8": 0.7529296875, + "0.9": 0.6888020833333334 + }, + "matched_cos_abs": { + "0.5": 0.892687201499939, + "0.6": 0.8885923027992249, + "0.7": 0.8832919001579285, + "0.8": 0.8800186514854431, + "0.9": 0.8724757432937622 + }, + "unique_frac_abs": { + "0.5": 0.8445638020833334, + "0.6": 0.87353515625, + "0.7": 0.9033203125, + "0.8": 0.9290364583333334, + "0.9": 0.9617513020833334 + }, + "dedup_dcs_abs": { + "0.5": 0.7539312969699192, + "0.6": 0.7762166160682682, + "0.7": 0.7978955152793787, + "0.8": 0.8175694112433121, + "0.9": 0.8391046821488999 + }, + "matched_frac_abs": { + "0.5": 0.7802734375, + "0.6": 0.7848307291666666, + "0.7": 0.7760416666666666, + "0.8": 0.751953125, + "0.9": 0.6884765625 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.8620738983154297, + "dcs_abs": 0.8638548851013184, + "frac_sign_flipped": 0.028645833333333332, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8867824077606201, + "0.6": 0.8854392170906067, + "0.7": 0.8801518082618713, + "0.8": 0.8756371140480042, + "0.9": 0.8707349896430969 + }, + "unique_frac_signed": { + "0.5": 0.8473307291666666, + "0.6": 0.87353515625, + "0.7": 0.90283203125, + "0.8": 0.9288736979166666, + "0.9": 0.9607747395833334 + }, + "dedup_dcs_signed": { + "0.5": 0.7513979841799785, + "0.6": 0.7734622848511208, + "0.7": 0.7946292448614258, + "0.8": 0.8133562841588475, + "0.9": 0.8365801829204429 + }, + "matched_frac_signed": { + "0.5": 0.7734375, + "0.6": 0.7809244791666666, + "0.7": 0.7770182291666666, + "0.8": 0.7513020833333334, + "0.9": 0.6901041666666666 + }, + "matched_cos_abs": { + "0.5": 0.8874406218528748, + "0.6": 0.8861882090568542, + "0.7": 0.8809804916381836, + "0.8": 0.8767107129096985, + "0.9": 0.8725345730781555 + }, + "unique_frac_abs": { + "0.5": 0.8458658854166666, + "0.6": 0.8723958333333334, + "0.7": 0.90185546875, + "0.8": 0.92822265625, + "0.9": 0.96044921875 + }, + "dedup_dcs_abs": { + "0.5": 0.7506557473582991, + "0.6": 0.7731069011303286, + "0.7": 0.7945170742459595, + "0.8": 0.8137827466998715, + "0.9": 0.8380251490452792 + }, + "matched_frac_abs": { + "0.5": 0.7737630208333334, + "0.6": 0.7822265625, + "0.7": 0.7760416666666666, + "0.8": 0.7506510416666666, + "0.9": 0.6888020833333334 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.8889439702033997, + "dcs_abs": 0.8901732563972473, + "frac_sign_flipped": 0.0185546875, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8958790898323059, + "0.6": 0.893430769443512, + "0.7": 0.889135479927063, + "0.8": 0.8864941000938416, + "0.9": 0.8865686058998108 + }, + "unique_frac_signed": { + "0.5": 0.8426106770833334, + "0.6": 0.86767578125, + "0.7": 0.8963216145833334, + "0.8": 0.9215494791666666, + "0.9": 0.9539388020833334 + }, + "dedup_dcs_signed": { + "0.5": 0.7548772864683997, + "0.6": 0.7752082408696879, + "0.7": 0.7969513489515521, + "0.8": 0.8169481762258025, + "0.9": 0.8457321938767564 + }, + "matched_frac_signed": { + "0.5": 0.7760416666666666, + "0.6": 0.7858072916666666, + "0.7": 0.78515625, + "0.8": 0.7705078125, + "0.9": 0.7272135416666666 + }, + "matched_cos_abs": { + "0.5": 0.8988944888114929, + "0.6": 0.8942874073982239, + "0.7": 0.8906096816062927, + "0.8": 0.8878304362297058, + "0.9": 0.8880638480186462 + }, + "unique_frac_abs": { + "0.5": 0.8411458333333334, + "0.6": 0.8660481770833334, + "0.7": 0.8953450520833334, + "0.8": 0.9207356770833334, + "0.9": 0.9534505208333334 + }, + "dedup_dcs_abs": { + "0.5": 0.7561013538700839, + "0.6": 0.7744959789658121, + "0.7": 0.7974029717637071, + "0.8": 0.8174571578371495, + "0.9": 0.8467249384266324 + }, + "matched_frac_abs": { + "0.5": 0.77734375, + "0.6": 0.787109375, + "0.7": 0.7854817708333334, + "0.8": 0.7708333333333334, + "0.9": 0.7259114583333334 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 0, + "n_dense_std": 0.0, + "l0_mean": null, + "l0_std": null, + "recon_r2_mean": null, + "recon_r2_std": null, + "l1_per_token_mean": null, + "l1_per_token_std": null, + "babel_mu1_signed_mean": 0.36736975113550824, + "babel_mu1_signed_std": 0.0028217504156440275, + "n_dup_signed_per_threshold": { + "0.5": 643.3333333333334, + "0.6": 528.6666666666666, + "0.7": 406.3333333333333, + "0.8": 295.0, + "0.9": 178.33333333333334 + }, + "n_dup_signed_band": 352.0833333333333, + "n_dup_signed_band_std": 27.894145144336893, + "n_dup_clusters_signed_per_threshold": { + "0.5": 169.33333333333334, + "0.6": 136.0, + "0.7": 103.0, + "0.8": 70.0, + "0.9": 52.0 + }, + "n_dup_clusters_signed_band": 90.25, + "n_dup_clusters_signed_band_std": 1.5206906325745548, + "largest_dup_signed_per_threshold": { + "0.5": 100.66666666666667, + "0.6": 53.666666666666664, + "0.7": 43.333333333333336, + "0.8": 31.333333333333332, + "0.9": 12.666666666666666 + }, + "largest_dup_signed_band": 35.25, + "largest_dup_signed_band_std": 8.011710179481033, + "dcs_signed_mean": 0.8715717395146688, + "dcs_signed_std": 0.015066675087434934, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.7535539138641777, + "0.6": 0.7748305659172022, + "0.7": 0.7962209682542986, + "0.8": 0.8155469306188428, + "0.9": 0.839661045659644 + }, + "dedup_dcs_signed_band": 0.8065648776124968, + "dedup_dcs_signed_band_std": 0.002102852872293367, + "matched_cos_signed_per_threshold": { + "0.5": 0.8910472989082336, + "0.6": 0.8883976737658182, + "0.7": 0.8834696213404337, + "0.8": 0.8800178567568461, + "0.9": 0.8757004737854004 + }, + "matched_cos_signed_band": 0.8818964064121246, + "matched_cos_signed_band_std": 0.006084734579468042, + "unique_frac_signed_per_threshold": { + "0.5": 0.845703125, + "0.6": 0.8721788194444445, + "0.7": 0.9012586805555557, + "0.8": 0.9267578125, + "0.9": 0.9588758680555557 + }, + "unique_frac_signed_band": 0.914767795138889, + "unique_frac_signed_band_std": 0.004299758816805034, + "matched_frac_signed_per_threshold": { + "0.5": 0.7765842013888888, + "0.6": 0.7835286458333334, + "0.7": 0.779079861111111, + "0.8": 0.7582465277777778, + "0.9": 0.7020399305555555 + }, + "matched_frac_signed_band": 0.7557237413194444, + "matched_frac_signed_band_std": 0.009915146747727952, + "babel_mu1_abs_mean": 0.37611472606658936, + "babel_mu1_abs_std": 0.003267168590536138, + "n_dup_abs_per_threshold": { + "0.5": 652.3333333333334, + "0.6": 535.6666666666666, + "0.7": 410.6666666666667, + "0.8": 299.6666666666667, + "0.9": 180.66666666666666 + }, + "n_dup_abs_band": 356.6666666666667, + "n_dup_abs_band_std": 28.583576286625387, + "n_dup_clusters_abs_per_threshold": { + "0.5": 172.66666666666666, + "0.6": 138.33333333333334, + "0.7": 104.0, + "0.8": 72.33333333333333, + "0.9": 53.333333333333336 + }, + "n_dup_clusters_abs_band": 92.0, + "n_dup_clusters_abs_band_std": 2.462214450449026, + "largest_dup_abs_per_threshold": { + "0.5": 80.33333333333333, + "0.6": 60.666666666666664, + "0.7": 45.0, + "0.8": 31.333333333333332, + "0.9": 11.666666666666666 + }, + "largest_dup_abs_band": 37.166666666666664, + "largest_dup_abs_band_std": 5.346338310781814, + "dcs_abs_mean": 0.8733628789583842, + "dcs_abs_std": 0.014599923677861167, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.7535627993994342, + "0.6": 0.7746064987214697, + "0.7": 0.7966051870963485, + "0.8": 0.8162697719267777, + "0.9": 0.8412849232069372 + }, + "dedup_dcs_abs_band": 0.8071915952378833, + "dedup_dcs_abs_band_std": 0.002126596326879785, + "matched_cos_abs_per_threshold": { + "0.5": 0.8930074373881022, + "0.6": 0.889689306418101, + "0.7": 0.8849606911341349, + "0.8": 0.8815199335416158, + "0.9": 0.877691388130188 + }, + "matched_cos_abs_band": 0.8834653298060099, + "matched_cos_abs_band_std": 0.005914915546760715, + "unique_frac_abs_per_threshold": { + "0.5": 0.8438585069444445, + "0.6": 0.8706597222222223, + "0.7": 0.9001736111111112, + "0.8": 0.925998263888889, + "0.9": 0.9585503472222223 + }, + "unique_frac_abs_band": 0.9138454861111112, + "unique_frac_abs_band_std": 0.00432777723906263, + "matched_frac_abs_per_threshold": { + "0.5": 0.7771267361111112, + "0.6": 0.7847222222222222, + "0.7": 0.7791883680555555, + "0.8": 0.7578125, + "0.9": 0.7010633680555557 + }, + "matched_frac_abs_band": 0.7556966145833334, + "matched_frac_abs_band_std": 0.01008819208703062, + "alive_source": "stats", + "frac_denominator": "d_sae", + "d_sae": 3072 + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_32/seed1/config.json b/vision/dinov2-b_layer10/mp/x4/l0_32/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..d39c416079930c68d7c93074e3b2b5646a514bb7 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_32/seed1/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "k_iters": 32, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 32, + "residual_thr": 0.05, + "max_iters": 307, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/code_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x4/l0_64/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..641ddb41a29bb4274dfab5a4c5ae30936d7cd404 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 94.26257976653696, + "recon_r2": 0.975821934346816, + "l1_per_token": 160.12352198927357 + }, + "seed1": { + "l0": 96.0824747081712, + "recon_r2": 0.9758979026055418, + "l1_per_token": 160.55607186960745 + }, + "seed2": { + "l0": 96.16374837872893, + "recon_r2": 0.9758934109302257, + "l1_per_token": 160.4021775571133 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.686996152798104, + "mcc_abs": 0.6870703592996316, + "act_iou": 0.25547694536790305, + "act_iou_n_pairs": 3072, + "input_r2": 0.9846564578689908, + "matched_mcc_signed": { + "0.6": 0.8233672574692353, + "0.7": 0.8235882716727134, + "0.8": 0.8303783596636519, + "0.9": 0.8476126278762411 + }, + "matched_mcc_abs": { + "0.6": 0.8221728981196199, + "0.7": 0.822716947685233, + "0.8": 0.831733380016577, + "0.9": 0.8478334868152545 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6846840780527694, + "mcc_abs": 0.6847622419526243, + "act_iou": 0.25246337038642785, + "act_iou_n_pairs": 3072, + "input_r2": 0.9844104815350083, + "matched_mcc_signed": { + "0.6": 0.8227527574304134, + "0.7": 0.8178582045636273, + "0.8": 0.8274592190546871, + "0.9": 0.844338840962079 + }, + "matched_mcc_abs": { + "0.6": 0.8202325549355041, + "0.7": 0.8191295136242774, + "0.8": 0.8274661268153322, + "0.9": 0.8450663089562113 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6946268447282802, + "mcc_abs": 0.6946844550087684, + "act_iou": 0.2577866682674802, + "act_iou_n_pairs": 3072, + "input_r2": 0.9846777056136713, + "matched_mcc_signed": { + "0.6": 0.822279493121979, + "0.7": 0.818058897473975, + "0.8": 0.8203782644254983, + "0.9": 0.8269563531697791 + }, + "matched_mcc_abs": { + "0.6": 0.820086935363402, + "0.7": 0.8187139526860329, + "0.8": 0.819773872370219, + "0.9": 0.8279184476094364 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 95.50293428447902, + "l0_std": 1.0749469056730019, + "recon_r2_mean": 0.975871082627528, + "recon_r2_std": 4.2622868386407093e-05, + "l1_per_token_mean": 160.3605904719981, + "l1_per_token_std": 0.2192531960553768, + "input_r2_mean": 0.9845815483392236, + "input_r2_std": 0.00014852863458536, + "mcc_signed_mean": 0.6887690251930513, + "mcc_signed_std": 0.005203072139369793, + "mcc_abs_mean": 0.6888390187536748, + "mcc_abs_std": 0.005192176343419167, + "matched_mcc_signed_per_threshold": { + "0.6": 0.8227998360072094, + "0.7": 0.8198351245701052, + "0.8": 0.8260719477146125, + "0.9": 0.8396359406693664 + }, + "matched_mcc_signed_band": 0.8270857122403233, + "matched_mcc_abs_per_threshold": { + "0.6": 0.8208307961395086, + "0.7": 0.8201868046651811, + "0.8": 0.8263244597340428, + "0.9": 0.8402727477936341 + }, + "matched_mcc_abs_band": 0.8269037020830917, + "act_iou_mean": 0.25524232800727037, + "act_iou_std": 0.002669393013783618, + "act_iou_n_pairs": 3072.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/dict_metrics_step_final.json b/vision/dinov2-b_layer10/mp/x4/l0_64/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..3cc4fa639964086adc9fb5a8440b0d7fe5b772b1 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/dict_metrics_step_final.json @@ -0,0 +1,595 @@ +{ + "schema": 3, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 633, + "0.6": 505, + "0.7": 383, + "0.8": 264, + "0.9": 129 + }, + "n_dup_clusters_signed": { + "0.5": 178, + "0.6": 140, + "0.7": 110, + "0.8": 71, + "0.9": 33 + }, + "n_dup_removed_signed": { + "0.5": 455, + "0.6": 365, + "0.7": 273, + "0.8": 193, + "0.9": 96 + }, + "largest_dup_signed": { + "0.5": 102, + "0.6": 56, + "0.7": 46, + "0.8": 27, + "0.9": 14 + }, + "babel_mu1_signed": 0.3663797080516815, + "n_dup_abs": { + "0.5": 644, + "0.6": 511, + "0.7": 389, + "0.8": 267, + "0.9": 130 + }, + "n_dup_clusters_abs": { + "0.5": 181, + "0.6": 142, + "0.7": 112, + "0.8": 72, + "0.9": 34 + }, + "n_dup_removed_abs": { + "0.5": 463, + "0.6": 369, + "0.7": 277, + "0.8": 195, + "0.9": 96 + }, + "largest_dup_abs": { + "0.5": 106, + "0.6": 55, + "0.7": 51, + "0.8": 27, + "0.9": 11 + }, + "babel_mu1_abs": 0.375279039144516, + "dup_excludes_dead": true + }, + "seed1": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 658, + "0.6": 541, + "0.7": 413, + "0.8": 303, + "0.9": 198 + }, + "n_dup_clusters_signed": { + "0.5": 174, + "0.6": 140, + "0.7": 100, + "0.8": 65, + "0.9": 60 + }, + "n_dup_removed_signed": { + "0.5": 484, + "0.6": 401, + "0.7": 313, + "0.8": 238, + "0.9": 138 + }, + "largest_dup_signed": { + "0.5": 89, + "0.6": 48, + "0.7": 28, + "0.8": 22, + "0.9": 10 + }, + "babel_mu1_signed": 0.3705531060695648, + "n_dup_abs": { + "0.5": 671, + "0.6": 548, + "0.7": 414, + "0.8": 308, + "0.9": 200 + }, + "n_dup_clusters_abs": { + "0.5": 179, + "0.6": 140, + "0.7": 97, + "0.8": 67, + "0.9": 61 + }, + "n_dup_removed_abs": { + "0.5": 492, + "0.6": 408, + "0.7": 317, + "0.8": 241, + "0.9": 139 + }, + "largest_dup_abs": { + "0.5": 73, + "0.6": 70, + "0.7": 28, + "0.8": 22, + "0.9": 10 + }, + "babel_mu1_abs": 0.3797185719013214, + "dup_excludes_dead": true + }, + "seed2": { + "d_sae": 3072, + "d_in": 768, + "l0": null, + "l0_raw": null, + "recon_r2": null, + "l1_per_token": null, + "stats_source": "mp-excluded", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 0, + "dense_rate": 0.0, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 639, + "0.6": 540, + "0.7": 423, + "0.8": 318, + "0.9": 208 + }, + "n_dup_clusters_signed": { + "0.5": 156, + "0.6": 128, + "0.7": 99, + "0.8": 74, + "0.9": 63 + }, + "n_dup_removed_signed": { + "0.5": 483, + "0.6": 412, + "0.7": 324, + "0.8": 244, + "0.9": 145 + }, + "largest_dup_signed": { + "0.5": 111, + "0.6": 57, + "0.7": 56, + "0.8": 45, + "0.9": 14 + }, + "babel_mu1_signed": 0.3651764392852783, + "n_dup_abs": { + "0.5": 642, + "0.6": 548, + "0.7": 429, + "0.8": 324, + "0.9": 212 + }, + "n_dup_clusters_abs": { + "0.5": 158, + "0.6": 133, + "0.7": 103, + "0.8": 78, + "0.9": 65 + }, + "n_dup_removed_abs": { + "0.5": 484, + "0.6": 415, + "0.7": 326, + "0.8": 246, + "0.9": 147 + }, + "largest_dup_abs": { + "0.5": 62, + "0.6": 57, + "0.7": 56, + "0.8": 45, + "0.9": 14 + }, + "babel_mu1_abs": 0.37334656715393066, + "dup_excludes_dead": true + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.863697350025177, + "dcs_abs": 0.8660604953765869, + "frac_sign_flipped": 0.033528645833333336, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8904803991317749, + "0.6": 0.8863230347633362, + "0.7": 0.8811215758323669, + "0.8": 0.8779223561286926, + "0.9": 0.8697978258132935 + }, + "unique_frac_signed": { + "0.5": 0.84716796875, + "0.6": 0.8753255208333334, + "0.7": 0.9046223958333334, + "0.8": 0.9298502604166666, + "0.9": 0.9619140625 + }, + "dedup_dcs_signed": { + "0.5": 0.754386470944155, + "0.6": 0.7758211720307979, + "0.7": 0.7970823109499179, + "0.8": 0.8163363314718783, + "0.9": 0.8366707601817325 + }, + "matched_frac_signed": { + "0.5": 0.7802734375, + "0.6": 0.7838541666666666, + "0.7": 0.7750651041666666, + "0.8": 0.7529296875, + "0.9": 0.6888020833333334 + }, + "matched_cos_abs": { + "0.5": 0.892687201499939, + "0.6": 0.8885923027992249, + "0.7": 0.8832919001579285, + "0.8": 0.8800186514854431, + "0.9": 0.8724757432937622 + }, + "unique_frac_abs": { + "0.5": 0.8445638020833334, + "0.6": 0.87353515625, + "0.7": 0.9033203125, + "0.8": 0.9290364583333334, + "0.9": 0.9617513020833334 + }, + "dedup_dcs_abs": { + "0.5": 0.7539312969699192, + "0.6": 0.7762166160682682, + "0.7": 0.7978955152793787, + "0.8": 0.8175694112433121, + "0.9": 0.8391046821488999 + }, + "matched_frac_abs": { + "0.5": 0.7802734375, + "0.6": 0.7848307291666666, + "0.7": 0.7760416666666666, + "0.8": 0.751953125, + "0.9": 0.6884765625 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.8620738983154297, + "dcs_abs": 0.8638548851013184, + "frac_sign_flipped": 0.028645833333333332, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8867824077606201, + "0.6": 0.8854392170906067, + "0.7": 0.8801518082618713, + "0.8": 0.8756371140480042, + "0.9": 0.8707349896430969 + }, + "unique_frac_signed": { + "0.5": 0.8473307291666666, + "0.6": 0.87353515625, + "0.7": 0.90283203125, + "0.8": 0.9288736979166666, + "0.9": 0.9607747395833334 + }, + "dedup_dcs_signed": { + "0.5": 0.7513979841799785, + "0.6": 0.7734622848511208, + "0.7": 0.7946292448614258, + "0.8": 0.8133562841588475, + "0.9": 0.8365801829204429 + }, + "matched_frac_signed": { + "0.5": 0.7734375, + "0.6": 0.7809244791666666, + "0.7": 0.7770182291666666, + "0.8": 0.7513020833333334, + "0.9": 0.6901041666666666 + }, + "matched_cos_abs": { + "0.5": 0.8874406218528748, + "0.6": 0.8861882090568542, + "0.7": 0.8809804916381836, + "0.8": 0.8767107129096985, + "0.9": 0.8725345730781555 + }, + "unique_frac_abs": { + "0.5": 0.8458658854166666, + "0.6": 0.8723958333333334, + "0.7": 0.90185546875, + "0.8": 0.92822265625, + "0.9": 0.96044921875 + }, + "dedup_dcs_abs": { + "0.5": 0.7506557473582991, + "0.6": 0.7731069011303286, + "0.7": 0.7945170742459595, + "0.8": 0.8137827466998715, + "0.9": 0.8380251490452792 + }, + "matched_frac_abs": { + "0.5": 0.7737630208333334, + "0.6": 0.7822265625, + "0.7": 0.7760416666666666, + "0.8": 0.7506510416666666, + "0.9": 0.6888020833333334 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.8889439702033997, + "dcs_abs": 0.8901732563972473, + "frac_sign_flipped": 0.0185546875, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8958790898323059, + "0.6": 0.893430769443512, + "0.7": 0.889135479927063, + "0.8": 0.8864941000938416, + "0.9": 0.8865686058998108 + }, + "unique_frac_signed": { + "0.5": 0.8426106770833334, + "0.6": 0.86767578125, + "0.7": 0.8963216145833334, + "0.8": 0.9215494791666666, + "0.9": 0.9539388020833334 + }, + "dedup_dcs_signed": { + "0.5": 0.7548772864683997, + "0.6": 0.7752082408696879, + "0.7": 0.7969513489515521, + "0.8": 0.8169481762258025, + "0.9": 0.8457321938767564 + }, + "matched_frac_signed": { + "0.5": 0.7760416666666666, + "0.6": 0.7858072916666666, + "0.7": 0.78515625, + "0.8": 0.7705078125, + "0.9": 0.7272135416666666 + }, + "matched_cos_abs": { + "0.5": 0.8988944888114929, + "0.6": 0.8942874073982239, + "0.7": 0.8906096816062927, + "0.8": 0.8878304362297058, + "0.9": 0.8880638480186462 + }, + "unique_frac_abs": { + "0.5": 0.8411458333333334, + "0.6": 0.8660481770833334, + "0.7": 0.8953450520833334, + "0.8": 0.9207356770833334, + "0.9": 0.9534505208333334 + }, + "dedup_dcs_abs": { + "0.5": 0.7561013538700839, + "0.6": 0.7744959789658121, + "0.7": 0.7974029717637071, + "0.8": 0.8174571578371495, + "0.9": 0.8467249384266324 + }, + "matched_frac_abs": { + "0.5": 0.77734375, + "0.6": 0.787109375, + "0.7": 0.7854817708333334, + "0.8": 0.7708333333333334, + "0.9": 0.7259114583333334 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 0, + "n_dense_std": 0.0, + "l0_mean": null, + "l0_std": null, + "recon_r2_mean": null, + "recon_r2_std": null, + "l1_per_token_mean": null, + "l1_per_token_std": null, + "babel_mu1_signed_mean": 0.36736975113550824, + "babel_mu1_signed_std": 0.0028217504156440275, + "n_dup_signed_per_threshold": { + "0.5": 643.3333333333334, + "0.6": 528.6666666666666, + "0.7": 406.3333333333333, + "0.8": 295.0, + "0.9": 178.33333333333334 + }, + "n_dup_signed_band": 352.0833333333333, + "n_dup_signed_band_std": 27.894145144336893, + "n_dup_clusters_signed_per_threshold": { + "0.5": 169.33333333333334, + "0.6": 136.0, + "0.7": 103.0, + "0.8": 70.0, + "0.9": 52.0 + }, + "n_dup_clusters_signed_band": 90.25, + "n_dup_clusters_signed_band_std": 1.5206906325745548, + "largest_dup_signed_per_threshold": { + "0.5": 100.66666666666667, + "0.6": 53.666666666666664, + "0.7": 43.333333333333336, + "0.8": 31.333333333333332, + "0.9": 12.666666666666666 + }, + "largest_dup_signed_band": 35.25, + "largest_dup_signed_band_std": 8.011710179481033, + "dcs_signed_mean": 0.8715717395146688, + "dcs_signed_std": 0.015066675087434934, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.7535539138641777, + "0.6": 0.7748305659172022, + "0.7": 0.7962209682542986, + "0.8": 0.8155469306188428, + "0.9": 0.839661045659644 + }, + "dedup_dcs_signed_band": 0.8065648776124968, + "dedup_dcs_signed_band_std": 0.002102852872293367, + "matched_cos_signed_per_threshold": { + "0.5": 0.8910472989082336, + "0.6": 0.8883976737658182, + "0.7": 0.8834696213404337, + "0.8": 0.8800178567568461, + "0.9": 0.8757004737854004 + }, + "matched_cos_signed_band": 0.8818964064121246, + "matched_cos_signed_band_std": 0.006084734579468042, + "unique_frac_signed_per_threshold": { + "0.5": 0.845703125, + "0.6": 0.8721788194444445, + "0.7": 0.9012586805555557, + "0.8": 0.9267578125, + "0.9": 0.9588758680555557 + }, + "unique_frac_signed_band": 0.914767795138889, + "unique_frac_signed_band_std": 0.004299758816805034, + "matched_frac_signed_per_threshold": { + "0.5": 0.7765842013888888, + "0.6": 0.7835286458333334, + "0.7": 0.779079861111111, + "0.8": 0.7582465277777778, + "0.9": 0.7020399305555555 + }, + "matched_frac_signed_band": 0.7557237413194444, + "matched_frac_signed_band_std": 0.009915146747727952, + "babel_mu1_abs_mean": 0.37611472606658936, + "babel_mu1_abs_std": 0.003267168590536138, + "n_dup_abs_per_threshold": { + "0.5": 652.3333333333334, + "0.6": 535.6666666666666, + "0.7": 410.6666666666667, + "0.8": 299.6666666666667, + "0.9": 180.66666666666666 + }, + "n_dup_abs_band": 356.6666666666667, + "n_dup_abs_band_std": 28.583576286625387, + "n_dup_clusters_abs_per_threshold": { + "0.5": 172.66666666666666, + "0.6": 138.33333333333334, + "0.7": 104.0, + "0.8": 72.33333333333333, + "0.9": 53.333333333333336 + }, + "n_dup_clusters_abs_band": 92.0, + "n_dup_clusters_abs_band_std": 2.462214450449026, + "largest_dup_abs_per_threshold": { + "0.5": 80.33333333333333, + "0.6": 60.666666666666664, + "0.7": 45.0, + "0.8": 31.333333333333332, + "0.9": 11.666666666666666 + }, + "largest_dup_abs_band": 37.166666666666664, + "largest_dup_abs_band_std": 5.346338310781814, + "dcs_abs_mean": 0.8733628789583842, + "dcs_abs_std": 0.014599923677861167, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.7535627993994342, + "0.6": 0.7746064987214697, + "0.7": 0.7966051870963485, + "0.8": 0.8162697719267777, + "0.9": 0.8412849232069372 + }, + "dedup_dcs_abs_band": 0.8071915952378833, + "dedup_dcs_abs_band_std": 0.002126596326879785, + "matched_cos_abs_per_threshold": { + "0.5": 0.8930074373881022, + "0.6": 0.889689306418101, + "0.7": 0.8849606911341349, + "0.8": 0.8815199335416158, + "0.9": 0.877691388130188 + }, + "matched_cos_abs_band": 0.8834653298060099, + "matched_cos_abs_band_std": 0.005914915546760715, + "unique_frac_abs_per_threshold": { + "0.5": 0.8438585069444445, + "0.6": 0.8706597222222223, + "0.7": 0.9001736111111112, + "0.8": 0.925998263888889, + "0.9": 0.9585503472222223 + }, + "unique_frac_abs_band": 0.9138454861111112, + "unique_frac_abs_band_std": 0.00432777723906263, + "matched_frac_abs_per_threshold": { + "0.5": 0.7771267361111112, + "0.6": 0.7847222222222222, + "0.7": 0.7791883680555555, + "0.8": 0.7578125, + "0.9": 0.7010633680555557 + }, + "matched_frac_abs_band": 0.7556966145833334, + "matched_frac_abs_band_std": 0.01008819208703062, + "alive_source": "stats", + "frac_denominator": "d_sae", + "d_sae": 3072 + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/seed0/config.json b/vision/dinov2-b_layer10/mp/x4/l0_64/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..e29bc53bb30892fe2d225d5ad62f012886441434 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/seed0/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "k_iters": 32, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 32, + "residual_thr": 0.05, + "max_iters": 307, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..c3acbb68a91b2c514b984518c5fcfed0d3c9f9a0 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.1779, + "r2_score": 0.9518, + "l0_per_token_0": 0.0077, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 93.5173 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/seed1/config.json b/vision/dinov2-b_layer10/mp/x4/l0_64/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..d39c416079930c68d7c93074e3b2b5646a514bb7 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/seed1/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "k_iters": 32, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 32, + "residual_thr": 0.05, + "max_iters": 307, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..79c066fddd1b730d9af6bbdb0134541233dabbec --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.1793, + "r2_score": 0.9515, + "l0_per_token_0": 0.0076, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 92.3942 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/seed2/config.json b/vision/dinov2-b_layer10/mp/x4/l0_64/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..bbfe8b7905bce01e56935c8b4e6059bb883ce1f9 --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/seed2/config.json @@ -0,0 +1,18 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "mp", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "k_iters": 32, + "mp_nonneg": true, + "mp_dropout": 0.0, + "inference_mode": "adaptive", + "inference_iters": 32, + "residual_thr": 0.05, + "max_iters": 307, + "patience": 2, + "class_type": "MPSAE", + "module_name": "consdl.models.mp" +} diff --git a/vision/dinov2-b_layer10/mp/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/mp/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..7918ca983198958a94e8cce82ec9731bbf9cd39b --- /dev/null +++ b/vision/dinov2-b_layer10/mp/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.179, + "r2_score": 0.9515, + "l0_per_token_0": 0.0077, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 92.3052 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_128/seed0/config.json b/vision/dinov2-b_layer10/relu/x16/l0_128/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..5949eb30b783de244a80f491fa25f7af3f2edec1 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_128/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..b55cdb595a9278a3e6b4c9814e758588ff1f6e19 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1393, + "r2_score": 0.9623, + "l0_per_token_0": 0.0101, + "l0_per_token_0.001": 0.0101, + "l0_per_token_0.01": 0.0098, + "l0_per_token_0.05": 0.0084, + "l0_per_token_0.1": 0.0071, + "l0_per_token_0.5": 0.0025, + "l0_per_token_1.0": 0.0012, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0001, + "live_latent_frac": 0.996, + "l1_per_token": 65.3288 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_128/seed1/config.json b/vision/dinov2-b_layer10/relu/x16/l0_128/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..4c15515c994745470445ff31f3f2293062c96289 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_128/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..34132c8509043d1d64116b3a14098b66944801f7 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1388, + "r2_score": 0.9624, + "l0_per_token_0": 0.0102, + "l0_per_token_0.001": 0.0102, + "l0_per_token_0.01": 0.0098, + "l0_per_token_0.05": 0.0085, + "l0_per_token_0.1": 0.0071, + "l0_per_token_0.5": 0.0025, + "l0_per_token_1.0": 0.0012, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0001, + "live_latent_frac": 0.9961, + "l1_per_token": 65.4799 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_128/seed2/config.json b/vision/dinov2-b_layer10/relu/x16/l0_128/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..43913a5027e3e99b474c9dfef9490cb3c811038a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_128/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..743ced9269a9dc43f418e1d1f1ee52b628d91112 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.1395, + "r2_score": 0.9639, + "l0_per_token_0": 0.0102, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 0.9976, + "l1_per_token": 65.1052 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_32/seed0/config.json b/vision/dinov2-b_layer10/relu/x16/l0_32/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..5949eb30b783de244a80f491fa25f7af3f2edec1 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_32/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..1bbc0667f6ed8f99c782671442a70fea3845cd95 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2285, + "r2_score": 0.9381, + "l0_per_token_0": 0.0027, + "l0_per_token_0.001": 0.0027, + "l0_per_token_0.01": 0.0027, + "l0_per_token_0.05": 0.0025, + "l0_per_token_0.1": 0.0024, + "l0_per_token_0.5": 0.0015, + "l0_per_token_1.0": 0.001, + "dead_rate": 0.0046, + "dense_rate_0.2": 0.0001, + "live_latent_frac": 0.7634, + "l1_per_token": 44.3201 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_32/seed1/config.json b/vision/dinov2-b_layer10/relu/x16/l0_32/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..4c15515c994745470445ff31f3f2293062c96289 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_32/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..dca5e3c4204e571c19a0080a11333f07a84e3edf --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.228, + "r2_score": 0.9382, + "l0_per_token_0": 0.0027, + "l0_per_token_0.001": 0.0027, + "l0_per_token_0.01": 0.0027, + "l0_per_token_0.05": 0.0025, + "l0_per_token_0.1": 0.0024, + "l0_per_token_0.5": 0.0015, + "l0_per_token_1.0": 0.001, + "dead_rate": 0.0098, + "dense_rate_0.2": 0.0001, + "live_latent_frac": 0.7194, + "l1_per_token": 44.3249 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_32/seed2/config.json b/vision/dinov2-b_layer10/relu/x16/l0_32/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..43913a5027e3e99b474c9dfef9490cb3c811038a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_32/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..c5e2d639946a06919c5e33ff78c113cf86f06ae4 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2272, + "r2_score": 0.9385, + "l0_per_token_0": 0.0027, + "l0_per_token_0.001": 0.0027, + "l0_per_token_0.01": 0.0027, + "l0_per_token_0.05": 0.0025, + "l0_per_token_0.1": 0.0024, + "l0_per_token_0.5": 0.0015, + "l0_per_token_1.0": 0.001, + "dead_rate": 0.0091, + "dense_rate_0.2": 0.0001, + "live_latent_frac": 0.7134, + "l1_per_token": 44.2982 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..9109d7ed191d2b5d4c03aa3f30c9b918b861fe1f --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2942, + "r2_score": 0.9203, + "l0_per_token_0": 0.0154, + "l0_per_token_0.001": 0.0152, + "l0_per_token_0.01": 0.014, + "l0_per_token_0.05": 0.0101, + "l0_per_token_0.1": 0.0071, + "l0_per_token_0.5": 0.0019, + "l0_per_token_1.0": 0.0009, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0007, + "live_latent_frac": 1.0, + "l1_per_token": 55.4259 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..37235b350f02b73e81faabef4ae0830a257e04f7 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2924, + "r2_score": 0.9208, + "l0_per_token_0": 0.0152, + "l0_per_token_0.001": 0.0151, + "l0_per_token_0.01": 0.0139, + "l0_per_token_0.05": 0.01, + "l0_per_token_0.1": 0.0071, + "l0_per_token_0.5": 0.0019, + "l0_per_token_1.0": 0.0009, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 1.0, + "l1_per_token": 55.5205 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x16/l0_64/seed2/config.json b/vision/dinov2-b_layer10/relu/x16/l0_64/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..43913a5027e3e99b474c9dfef9490cb3c811038a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_64/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 16, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..02dad6deec1e35f5e3612867bdcd7331787dc99c --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.1815, + "r2_score": 0.9509, + "l0_per_token_0": 0.0051, + "dead_rate": 0.0041, + "dense_rate_0.2": 0.0001, + "live_latent_frac": 0.9136, + "l1_per_token": 52.361 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/code_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x4/l0_128/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..9ab0f9aac5e5a87473a91391f8aa8ec8c7420d0a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 122.72297276264591, + "recon_r2": 0.9616755973879049, + "l1_per_token": 82.92275740473966 + }, + "seed1": { + "l0": 122.41150972762645, + "recon_r2": 0.961533296011064, + "l1_per_token": 82.74821288792349 + }, + "seed2": { + "l0": 123.16197924773022, + "recon_r2": 0.9615754623232561, + "l1_per_token": 82.97770552388067 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.8374067763608704, + "mcc_abs": 0.8374269454509773, + "act_iou": 0.5933373169240851, + "act_iou_n_pairs": 3072, + "input_r2": 0.9944742545073759, + "matched_mcc_signed": { + "0.6": 0.9527467247067329, + "0.7": 0.9656721903191003, + "0.8": 0.9743078537020083, + "0.9": 0.9840360988016991 + }, + "matched_mcc_abs": { + "0.6": 0.9530619326262914, + "0.7": 0.9654634140177693, + "0.8": 0.9744039048154909, + "0.9": 0.9840360988016991 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.8366777678134731, + "mcc_abs": 0.836683817644106, + "act_iou": 0.5932644984775887, + "act_iou_n_pairs": 3072, + "input_r2": 0.9944222537724693, + "matched_mcc_signed": { + "0.6": 0.9514477924077283, + "0.7": 0.9648179867009476, + "0.8": 0.9744669183015273, + "0.9": 0.9841673144447168 + }, + "matched_mcc_abs": { + "0.6": 0.951596527860768, + "0.7": 0.9647453158008812, + "0.8": 0.9749625859248909, + "0.9": 0.9841673144447168 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.8311535658614195, + "mcc_abs": 0.8311755591749804, + "act_iou": 0.5839698936071955, + "act_iou_n_pairs": 3072, + "input_r2": 0.9942746075270777, + "matched_mcc_signed": { + "0.6": 0.9528390309951118, + "0.7": 0.9642436437791675, + "0.8": 0.9733988392457352, + "0.9": 0.9848346523410654 + }, + "matched_mcc_abs": { + "0.6": 0.9523186649763756, + "0.7": 0.9645176640441809, + "0.8": 0.9732356899366137, + "0.9": 0.9848346523410654 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 122.76548724600086, + "l0_std": 0.37703678351375347, + "recon_r2_mean": 0.9615947852407416, + "recon_r2_std": 7.309207766796224e-05, + "l1_per_token_mean": 82.88289193884793, + "l1_per_token_std": 0.11982762209823392, + "input_r2_mean": 0.9943903719356411, + "input_r2_std": 0.00010357155895903448, + "mcc_signed_mean": 0.835079370011921, + "mcc_signed_std": 0.003419329910467184, + "mcc_abs_mean": 0.835095440756688, + "mcc_abs_std": 0.003414990987585991, + "matched_mcc_signed_per_threshold": { + "0.6": 0.9523445160365244, + "0.7": 0.9649112735997386, + "0.8": 0.9740578704164236, + "0.9": 0.9843460218624939 + }, + "matched_mcc_signed_band": 0.9689149204787951, + "matched_mcc_abs_per_threshold": { + "0.6": 0.9523257084878116, + "0.7": 0.9649087979542771, + "0.8": 0.9742007268923318, + "0.9": 0.9843460218624939 + }, + "matched_mcc_abs_band": 0.9689453137992287, + "act_iou_mean": 0.5901905696696231, + "act_iou_std": 0.005387386531221052, + "act_iou_n_pairs": 3072.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/dict_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x4/l0_128/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..517963e8b97946887a2521e4057ea34a2bdc42be --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/dict_metrics_step_final.json @@ -0,0 +1,545 @@ +{ + "schema": 2, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 3072, + "d_in": 768, + "l0": 123, + "l0_raw": 122.88, + "recon_r2": 0.9616, + "l1_per_token": 82.9586, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 250, + "dense_rate": 0.08138020833333333, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 33, + "0.6": 9, + "0.7": 8, + "0.8": 6, + "0.9": 5 + }, + "n_dup_clusters_signed": { + "0.5": 14, + "0.6": 2, + "0.7": 3, + "0.8": 2, + "0.9": 2 + }, + "largest_dup_signed": { + "0.5": 6, + "0.6": 6, + "0.7": 4, + "0.8": 4, + "0.9": 3 + }, + "babel_mu1_signed": 0.2591010332107544, + "n_dup_abs": { + "0.5": 126, + "0.6": 39, + "0.7": 16, + "0.8": 6, + "0.9": 5 + }, + "n_dup_clusters_abs": { + "0.5": 60, + "0.6": 17, + "0.7": 7, + "0.8": 2, + "0.9": 2 + }, + "largest_dup_abs": { + "0.5": 6, + "0.6": 6, + "0.7": 4, + "0.8": 4, + "0.9": 3 + }, + "babel_mu1_abs": 0.2791454493999481 + }, + "seed1": { + "d_sae": 3072, + "d_in": 768, + "l0": 123, + "l0_raw": 122.5728, + "recon_r2": 0.9615, + "l1_per_token": 82.7696, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 249, + "dense_rate": 0.0810546875, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 36, + "0.6": 12, + "0.7": 9, + "0.8": 6, + "0.9": 4 + }, + "n_dup_clusters_signed": { + "0.5": 16, + "0.6": 4, + "0.7": 3, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_signed": { + "0.5": 5, + "0.6": 5, + "0.7": 5, + "0.8": 4, + "0.9": 4 + }, + "babel_mu1_signed": 0.25992587208747864, + "n_dup_abs": { + "0.5": 119, + "0.6": 53, + "0.7": 16, + "0.8": 6, + "0.9": 4 + }, + "n_dup_clusters_abs": { + "0.5": 57, + "0.6": 24, + "0.7": 6, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_abs": { + "0.5": 6, + "0.6": 6, + "0.7": 6, + "0.8": 4, + "0.9": 4 + }, + "babel_mu1_abs": 0.278845876455307 + }, + "seed2": { + "d_sae": 3072, + "d_in": 768, + "l0": 123, + "l0_raw": 123.18719999999999, + "recon_r2": 0.9615, + "l1_per_token": 82.9934, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 242, + "dense_rate": 0.07877604166666667, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 33, + "0.6": 10, + "0.7": 8, + "0.8": 5, + "0.9": 3 + }, + "n_dup_clusters_signed": { + "0.5": 15, + "0.6": 4, + "0.7": 3, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_signed": { + "0.5": 4, + "0.6": 4, + "0.7": 4, + "0.8": 3, + "0.9": 3 + }, + "babel_mu1_signed": 0.2590630352497101, + "n_dup_abs": { + "0.5": 102, + "0.6": 35, + "0.7": 8, + "0.8": 5, + "0.9": 3 + }, + "n_dup_clusters_abs": { + "0.5": 49, + "0.6": 16, + "0.7": 3, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_abs": { + "0.5": 5, + "0.6": 5, + "0.7": 4, + "0.8": 3, + "0.9": 3 + }, + "babel_mu1_abs": 0.2777848243713379 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.8615548014640808, + "dcs_abs": 0.8619464039802551, + "frac_sign_flipped": 0.015299479166666666, + "alive_source": "stats", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8615650534629822, + "0.6": 0.8615062832832336, + "0.7": 0.8615275025367737, + "0.8": 0.8613839745521545, + "0.9": 0.8614097237586975 + }, + "unique_frac_signed": { + "0.5": 0.99365234375, + "0.6": 0.99755859375, + "0.7": 0.9982096354166666, + "0.8": 0.9986979166666666, + "0.9": 0.9990234375 + }, + "dedup_dcs_signed": { + "0.5": 0.8560961346665863, + "0.6": 0.8594029964588117, + "0.7": 0.8599850542086642, + "0.8": 0.8602623808352897, + "0.9": 0.8605685033253394 + }, + "matched_frac_signed": { + "0.5": 0.8583984375, + "0.6": 0.8375651041666666, + "0.7": 0.8128255208333334, + "0.8": 0.7893880208333334, + "0.9": 0.7421875 + }, + "matched_cos_abs": { + "0.5": 0.8678025007247925, + "0.6": 0.8645798563957214, + "0.7": 0.8624374270439148, + "0.8": 0.8617727160453796, + "0.9": 0.8618087768554688 + }, + "unique_frac_abs": { + "0.5": 0.9791666666666666, + "0.6": 0.99169921875, + "0.7": 0.9969075520833334, + "0.8": 0.9986979166666666, + "0.9": 0.9990234375 + }, + "dedup_dcs_abs": { + "0.5": 0.8497232819596926, + "0.6": 0.8574031681346241, + "0.7": 0.8597703842193976, + "0.8": 0.8606506161546955, + "0.9": 0.8609671667218208 + }, + "matched_frac_abs": { + "0.5": 0.8541666666666666, + "0.6": 0.8349609375, + "0.7": 0.8131510416666666, + "0.8": 0.7890625, + "0.9": 0.7421875 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.8606703281402588, + "dcs_abs": 0.8611266613006592, + "frac_sign_flipped": 0.015950520833333332, + "alive_source": "stats", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8607785105705261, + "0.6": 0.8606432676315308, + "0.7": 0.8604977130889893, + "0.8": 0.8605619072914124, + "0.9": 0.8605973720550537 + }, + "unique_frac_signed": { + "0.5": 0.9939778645833334, + "0.6": 0.9978841145833334, + "0.7": 0.9983723958333334, + "0.8": 0.9988606770833334, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_signed": { + "0.5": 0.8555947858161138, + "0.6": 0.8588222450925969, + "0.7": 0.8590971634257585, + "0.8": 0.8595814493892249, + "0.9": 0.8598970161207641 + }, + "matched_frac_signed": { + "0.5": 0.8561197916666666, + "0.6": 0.841796875, + "0.7": 0.8157552083333334, + "0.8": 0.7864583333333334, + "0.9": 0.740234375 + }, + "matched_cos_abs": { + "0.5": 0.8675013184547424, + "0.6": 0.8631106615066528, + "0.7": 0.8618268370628357, + "0.8": 0.8610517382621765, + "0.9": 0.8610787391662598 + }, + "unique_frac_abs": { + "0.5": 0.9806315104166666, + "0.6": 0.9933268229166666, + "0.7": 0.9977213541666666, + "0.8": 0.9988606770833334, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_abs": { + "0.5": 0.8506991282047238, + "0.6": 0.8573509712199059, + "0.7": 0.8598630389315076, + "0.8": 0.8600707222843388, + "0.9": 0.8603779914944122 + }, + "matched_frac_abs": { + "0.5": 0.8492838541666666, + "0.6": 0.83984375, + "0.7": 0.8157552083333334, + "0.8": 0.7861328125, + "0.9": 0.740234375 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.8549737930297852, + "dcs_abs": 0.8555158972740173, + "frac_sign_flipped": 0.013997395833333334, + "alive_source": "stats", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8556388020515442, + "0.6": 0.8550536036491394, + "0.7": 0.8549681901931763, + "0.8": 0.8550165891647339, + "0.9": 0.8550625443458557 + }, + "unique_frac_signed": { + "0.5": 0.9938151041666666, + "0.6": 0.9977213541666666, + "0.7": 0.9982096354166666, + "0.8": 0.9988606770833334, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_signed": { + "0.5": 0.8503467651898973, + "0.6": 0.8531052393179076, + "0.7": 0.8534374854255777, + "0.8": 0.8540424491705684, + "0.9": 0.8543666926658867 + }, + "matched_frac_signed": { + "0.5": 0.8502604166666666, + "0.6": 0.8294270833333334, + "0.7": 0.8063151041666666, + "0.8": 0.7789713541666666, + "0.9": 0.7252604166666666 + }, + "matched_cos_abs": { + "0.5": 0.8617808818817139, + "0.6": 0.8580303192138672, + "0.7": 0.8561945557594299, + "0.8": 0.8555225729942322, + "0.9": 0.8555683493614197 + }, + "unique_frac_abs": { + "0.5": 0.9812825520833334, + "0.6": 0.9921875, + "0.7": 0.99755859375, + "0.8": 0.9988606770833334, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_abs": { + "0.5": 0.8456505431095138, + "0.6": 0.8513269573450089, + "0.7": 0.8541042370197829, + "0.8": 0.8545478565210942, + "0.9": 0.8548720860562753 + }, + "matched_frac_abs": { + "0.5": 0.8463541666666666, + "0.6": 0.8274739583333334, + "0.7": 0.8059895833333334, + "0.8": 0.7796223958333334, + "0.9": 0.7252604166666666 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 247, + "n_dense_std": 4.163331998932266, + "l0_mean": 123, + "l0_std": 0.3071999999999946, + "recon_r2_mean": 0.9615333333333332, + "recon_r2_std": 5.773502691902032e-05, + "l1_per_token_mean": 82.9072, + "l1_per_token_std": 0.1203144768235883, + "babel_mu1_signed_mean": 0.259363313515981, + "babel_mu1_signed_std": 0.00048756032553412906, + "n_dup_signed_per_threshold": { + "0.5": 34.0, + "0.6": 10.333333333333334, + "0.7": 8.333333333333334, + "0.8": 5.666666666666667, + "0.9": 4.0 + }, + "n_dup_signed_band": 7.083333333333334, + "n_dup_signed_band_std": 0.6291528696058959, + "n_dup_clusters_signed_per_threshold": { + "0.5": 15.0, + "0.6": 3.3333333333333335, + "0.7": 3.0, + "0.8": 2.0, + "0.9": 1.3333333333333333 + }, + "n_dup_clusters_signed_band": 2.416666666666667, + "n_dup_clusters_signed_band_std": 0.14433756729740643, + "largest_dup_signed_per_threshold": { + "0.5": 5.0, + "0.6": 5.0, + "0.7": 4.333333333333333, + "0.8": 3.6666666666666665, + "0.9": 3.3333333333333335 + }, + "largest_dup_signed_band": 4.083333333333333, + "largest_dup_signed_band_std": 0.5204164998665333, + "dcs_signed_mean": 0.8590663075447083, + "dcs_signed_std": 0.0035717054128637706, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.8540125618908658, + "0.6": 0.857110160289772, + "0.7": 0.8575065676866668, + "0.8": 0.8579620931316944, + "0.9": 0.85827740403733 + }, + "dedup_dcs_signed_band": 0.8577140562863658, + "dedup_dcs_signed_band_std": 0.003461403816052478, + "matched_cos_signed_per_threshold": { + "0.5": 0.8593274553616842, + "0.6": 0.8590677181879679, + "0.7": 0.8589978019396464, + "0.8": 0.8589874903361002, + "0.9": 0.8590232133865356 + }, + "matched_cos_signed_band": 0.8590190559625626, + "matched_cos_signed_band_std": 0.003486741862333244, + "unique_frac_signed_per_threshold": { + "0.5": 0.9938151041666666, + "0.6": 0.9977213541666666, + "0.7": 0.9982638888888888, + "0.8": 0.9988064236111112, + "0.9": 0.9991319444444443 + }, + "unique_frac_signed_band": 0.9984809027777777, + "unique_frac_signed_band_std": 0.00010240118320410176, + "matched_frac_signed_per_threshold": { + "0.5": 0.8549262152777777, + "0.6": 0.8362630208333334, + "0.7": 0.8116319444444445, + "0.8": 0.7849392361111112, + "0.9": 0.7358940972222222 + }, + "matched_frac_signed_band": 0.7921820746527779, + "matched_frac_signed_band_std": 0.006232009709536583, + "babel_mu1_abs_mean": 0.278592050075531, + "babel_mu1_abs_std": 0.0007149447464124693, + "n_dup_abs_per_threshold": { + "0.5": 115.66666666666667, + "0.6": 42.333333333333336, + "0.7": 13.333333333333334, + "0.8": 5.666666666666667, + "0.9": 4.0 + }, + "n_dup_abs_band": 16.333333333333336, + "n_dup_abs_band_std": 3.5029749261639505, + "n_dup_clusters_abs_per_threshold": { + "0.5": 55.333333333333336, + "0.6": 19.0, + "0.7": 5.333333333333333, + "0.8": 2.0, + "0.9": 1.3333333333333333 + }, + "n_dup_clusters_abs_band": 6.916666666666666, + "n_dup_clusters_abs_band_std": 1.3768926368215255, + "largest_dup_abs_per_threshold": { + "0.5": 5.666666666666667, + "0.6": 5.666666666666667, + "0.7": 4.666666666666667, + "0.8": 3.6666666666666665, + "0.9": 3.3333333333333335 + }, + "largest_dup_abs_band": 4.333333333333333, + "largest_dup_abs_band_std": 0.6291528696058959, + "dcs_abs_mean": 0.8595296541849772, + "dcs_abs_std": 0.003500096844431084, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.8486909844246434, + "0.6": 0.8553603655665128, + "0.7": 0.8579125533902294, + "0.8": 0.8584230649867095, + "0.9": 0.8587390814241694 + }, + "dedup_dcs_abs_band": 0.8576087663419053, + "dedup_dcs_abs_band_std": 0.0033769675722479494, + "matched_cos_abs_per_threshold": { + "0.5": 0.8656949003537496, + "0.6": 0.8619069457054138, + "0.7": 0.8601529399553934, + "0.8": 0.8594490091005961, + "0.9": 0.8594852884610494 + }, + "matched_cos_abs_band": 0.8602485458056132, + "matched_cos_abs_band_std": 0.0034230421040169984, + "unique_frac_abs_per_threshold": { + "0.5": 0.9803602430555555, + "0.6": 0.9924045138888888, + "0.7": 0.9973958333333334, + "0.8": 0.9988064236111112, + "0.9": 0.9991319444444443 + }, + "unique_frac_abs_band": 0.9969346788194444, + "unique_frac_abs_band_std": 0.0003460652892299191, + "matched_frac_abs_per_threshold": { + "0.5": 0.8499348958333334, + "0.6": 0.8340928819444445, + "0.7": 0.8116319444444445, + "0.8": 0.7849392361111112, + "0.9": 0.7358940972222222 + }, + "matched_frac_abs_band": 0.7916395399305556, + "matched_frac_abs_band_std": 0.006116703056668404, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/seed0/config.json b/vision/dinov2-b_layer10/relu/x4/l0_128/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..920481756aae71a20da4a01e54beb062205c045a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..b90dab2f8a8db259f5dc2bcf1ed310268d8f1ef1 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1417, + "r2_score": 0.9616, + "l0_per_token_0": 0.04, + "l0_per_token_0.001": 0.0399, + "l0_per_token_0.01": 0.039, + "l0_per_token_0.05": 0.0354, + "l0_per_token_0.1": 0.0314, + "l0_per_token_0.5": 0.0136, + "l0_per_token_1.0": 0.0066, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0007, + "live_latent_frac": 1.0, + "l1_per_token": 82.9586 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/seed1/config.json b/vision/dinov2-b_layer10/relu/x4/l0_128/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2b729114fa7a39fd6d4550e7ebd1e771051ed38a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..44d568b9142b4bd0d09024d9a2614a7f680e8e3f --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1422, + "r2_score": 0.9615, + "l0_per_token_0": 0.0399, + "l0_per_token_0.001": 0.0398, + "l0_per_token_0.01": 0.0389, + "l0_per_token_0.05": 0.0353, + "l0_per_token_0.1": 0.0314, + "l0_per_token_0.5": 0.0135, + "l0_per_token_1.0": 0.0066, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0007, + "live_latent_frac": 1.0, + "l1_per_token": 82.7696 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/seed2/config.json b/vision/dinov2-b_layer10/relu/x4/l0_128/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..d2ac614f301c676dc8383e32f921621c96254746 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..274a0bab8e4678fd51c9843bea17b1a854244e5e --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.142, + "r2_score": 0.9615, + "l0_per_token_0": 0.0401, + "l0_per_token_0.001": 0.04, + "l0_per_token_0.01": 0.0391, + "l0_per_token_0.05": 0.0355, + "l0_per_token_0.1": 0.0315, + "l0_per_token_0.5": 0.0136, + "l0_per_token_1.0": 0.0066, + "dead_rate": 0.0, + "dense_rate_0.2": 0.001, + "live_latent_frac": 1.0, + "l1_per_token": 82.9934 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/code_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x4/l0_32/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..38b332ec0c64fc18c4a89887ccddefc4ae4cfc2f --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 28.75742671854734, + "recon_r2": 0.9331260497753574, + "l1_per_token": 43.41207277316912 + }, + "seed1": { + "l0": 28.789752269779505, + "recon_r2": 0.9332015297469484, + "l1_per_token": 43.37717555687601 + }, + "seed2": { + "l0": 28.86615175097276, + "recon_r2": 0.9331878685913503, + "l1_per_token": 43.476472826381595 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.7485656243303157, + "mcc_abs": 0.7485823145595981, + "act_iou": 0.5373324759634601, + "act_iou_n_pairs": 3072, + "input_r2": 0.9940921474898794, + "matched_mcc_signed": { + "0.6": 0.9362052948502222, + "0.7": 0.9468101560074615, + "0.8": 0.9585383132917017, + "0.9": 0.9746442610957369 + }, + "matched_mcc_abs": { + "0.6": 0.9360045344703435, + "0.7": 0.9465374585258786, + "0.8": 0.9585594597296806, + "0.9": 0.974644261095737 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.7527460151650014, + "mcc_abs": 0.7527653220790221, + "act_iou": 0.5370893013303905, + "act_iou_n_pairs": 3072, + "input_r2": 0.9940568967408995, + "matched_mcc_signed": { + "0.6": 0.9297467231365509, + "0.7": 0.945819039715201, + "0.8": 0.9577155841766599, + "0.9": 0.9730188822568914 + }, + "matched_mcc_abs": { + "0.6": 0.9306900226419536, + "0.7": 0.9440951684439712, + "0.8": 0.9567875631068614, + "0.9": 0.9729289378681873 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.7563688206662168, + "mcc_abs": 0.7563972732672634, + "act_iou": 0.5462570206108346, + "act_iou_n_pairs": 3072, + "input_r2": 0.9942806562448961, + "matched_mcc_signed": { + "0.6": 0.933769216982659, + "0.7": 0.9456652544347601, + "0.8": 0.9594729979192922, + "0.9": 0.9747398173239761 + }, + "matched_mcc_abs": { + "0.6": 0.9336644408760812, + "0.7": 0.9456143077045632, + "0.8": 0.9592330639268538, + "0.9": 0.9747398173239761 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 28.804443579766538, + "l0_std": 0.05583152434627349, + "recon_r2_mean": 0.9331718160378853, + "recon_r2_std": 4.021902388673153e-05, + "l1_per_token_mean": 43.42190705214224, + "l1_per_token_std": 0.050373819738104286, + "input_r2_mean": 0.9941432334918917, + "input_r2_std": 0.0001203096571701847, + "mcc_signed_mean": 0.752560153387178, + "mcc_signed_std": 0.003904916992011077, + "mcc_abs_mean": 0.7525816366352945, + "mcc_abs_std": 0.003910716067078541, + "matched_mcc_signed_per_threshold": { + "0.6": 0.9332404116564774, + "0.7": 0.9460981500524742, + "0.8": 0.9585756317958847, + "0.9": 0.9741343202255348 + }, + "matched_mcc_signed_band": 0.9530121284325928, + "matched_mcc_abs_per_threshold": { + "0.6": 0.9334529993294595, + "0.7": 0.945415644891471, + "0.8": 0.9581933622544653, + "0.9": 0.9741043387626335 + }, + "matched_mcc_abs_band": 0.9527915863095073, + "act_iou_mean": 0.5402262659682284, + "act_iou_std": 0.005224201818941274, + "act_iou_n_pairs": 3072.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/dict_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x4/l0_32/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..8de93f7ccc56a34ca6392d8962f9a3562b064e5c --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/dict_metrics_step_final.json @@ -0,0 +1,595 @@ +{ + "schema": 3, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 3072, + "d_in": 768, + "l0": 29, + "l0_raw": 28.876800000000003, + "recon_r2": 0.9331, + "l1_per_token": 43.4364, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 4, + "dense_rate": 0.0013020833333333333, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 238, + "0.6": 62, + "0.7": 18, + "0.8": 8, + "0.9": 5 + }, + "n_dup_clusters_signed": { + "0.5": 114, + "0.6": 28, + "0.7": 7, + "0.8": 3, + "0.9": 2 + }, + "n_dup_removed_signed": { + "0.5": 124, + "0.6": 34, + "0.7": 11, + "0.8": 5, + "0.9": 3 + }, + "largest_dup_signed": { + "0.5": 6, + "0.6": 5, + "0.7": 5, + "0.8": 4, + "0.9": 3 + }, + "babel_mu1_signed": 0.3512091636657715, + "n_dup_abs": { + "0.5": 241, + "0.6": 66, + "0.7": 20, + "0.8": 10, + "0.9": 5 + }, + "n_dup_clusters_abs": { + "0.5": 115, + "0.6": 29, + "0.7": 7, + "0.8": 4, + "0.9": 2 + }, + "n_dup_removed_abs": { + "0.5": 126, + "0.6": 37, + "0.7": 13, + "0.8": 6, + "0.9": 3 + }, + "largest_dup_abs": { + "0.5": 7, + "0.6": 6, + "0.7": 6, + "0.8": 4, + "0.9": 3 + }, + "babel_mu1_abs": 0.35238614678382874, + "dup_excludes_dead": true + }, + "seed1": { + "d_sae": 3072, + "d_in": 768, + "l0": 29, + "l0_raw": 28.876800000000003, + "recon_r2": 0.9332, + "l1_per_token": 43.4055, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 7, + "dense_rate": 0.0022786458333333335, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 258, + "0.6": 71, + "0.7": 22, + "0.8": 7, + "0.9": 6 + }, + "n_dup_clusters_signed": { + "0.5": 120, + "0.6": 31, + "0.7": 8, + "0.8": 2, + "0.9": 2 + }, + "n_dup_removed_signed": { + "0.5": 138, + "0.6": 40, + "0.7": 14, + "0.8": 5, + "0.9": 4 + }, + "largest_dup_signed": { + "0.5": 8, + "0.6": 7, + "0.7": 7, + "0.8": 5, + "0.9": 4 + }, + "babel_mu1_signed": 0.3531588017940521, + "n_dup_abs": { + "0.5": 261, + "0.6": 73, + "0.7": 24, + "0.8": 9, + "0.9": 6 + }, + "n_dup_clusters_abs": { + "0.5": 121, + "0.6": 31, + "0.7": 9, + "0.8": 3, + "0.9": 2 + }, + "n_dup_removed_abs": { + "0.5": 140, + "0.6": 42, + "0.7": 15, + "0.8": 6, + "0.9": 4 + }, + "largest_dup_abs": { + "0.5": 8, + "0.6": 8, + "0.7": 7, + "0.8": 5, + "0.9": 4 + }, + "babel_mu1_abs": 0.3542725741863251, + "dup_excludes_dead": true + }, + "seed2": { + "d_sae": 3072, + "d_in": 768, + "l0": 29, + "l0_raw": 28.876800000000003, + "recon_r2": 0.9362, + "l1_per_token": 43.494, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 5, + "dense_rate": 0.0016276041666666667, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 270, + "0.6": 67, + "0.7": 22, + "0.8": 15, + "0.9": 10 + }, + "n_dup_clusters_signed": { + "0.5": 127, + "0.6": 29, + "0.7": 8, + "0.8": 6, + "0.9": 4 + }, + "n_dup_removed_signed": { + "0.5": 143, + "0.6": 38, + "0.7": 14, + "0.8": 9, + "0.9": 6 + }, + "largest_dup_signed": { + "0.5": 7, + "0.6": 7, + "0.7": 6, + "0.8": 5, + "0.9": 4 + }, + "babel_mu1_signed": 0.35442742705345154, + "n_dup_abs": { + "0.5": 276, + "0.6": 70, + "0.7": 23, + "0.8": 15, + "0.9": 10 + }, + "n_dup_clusters_abs": { + "0.5": 128, + "0.6": 30, + "0.7": 9, + "0.8": 6, + "0.9": 4 + }, + "n_dup_removed_abs": { + "0.5": 148, + "0.6": 40, + "0.7": 14, + "0.8": 9, + "0.9": 6 + }, + "largest_dup_abs": { + "0.5": 9, + "0.6": 6, + "0.7": 5, + "0.8": 5, + "0.9": 4 + }, + "babel_mu1_abs": 0.35553470253944397, + "dup_excludes_dead": true + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.7903321385383606, + "dcs_abs": 0.7910934090614319, + "frac_sign_flipped": 0.0458984375, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.7921655774116516, + "0.6": 0.7915148735046387, + "0.7": 0.7910513877868652, + "0.8": 0.7903655767440796, + "0.9": 0.7904613614082336 + }, + "unique_frac_signed": { + "0.5": 0.9573567708333334, + "0.6": 0.9879557291666666, + "0.7": 0.9959309895833334, + "0.8": 0.9983723958333334, + "0.9": 0.9988606770833334 + }, + "dedup_dcs_signed": { + "0.5": 0.7583850791561417, + "0.6": 0.7819816539995372, + "0.7": 0.7878325914498419, + "0.8": 0.7890791744381811, + "0.9": 0.7895607706644417 + }, + "matched_frac_signed": { + "0.5": 0.763671875, + "0.6": 0.7692057291666666, + "0.7": 0.7545572916666666, + "0.8": 0.7272135416666666, + "0.9": 0.66796875 + }, + "matched_cos_abs": { + "0.5": 0.7930892109870911, + "0.6": 0.7921311259269714, + "0.7": 0.7917256951332092, + "0.8": 0.7910783886909485, + "0.9": 0.7912248969078064 + }, + "unique_frac_abs": { + "0.5": 0.9567057291666666, + "0.6": 0.9871419270833334, + "0.7": 0.9954427083333334, + "0.8": 0.998046875, + "0.9": 0.9988606770833334 + }, + "dedup_dcs_abs": { + "0.5": 0.7587529918916213, + "0.6": 0.7819458461502412, + "0.7": 0.7881175702204928, + "0.8": 0.7895333137130365, + "0.9": 0.7903234362505221 + }, + "matched_frac_abs": { + "0.5": 0.763671875, + "0.6": 0.7688802083333334, + "0.7": 0.7542317708333334, + "0.8": 0.7268880208333334, + "0.9": 0.66796875 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.7953314781188965, + "dcs_abs": 0.7961196899414062, + "frac_sign_flipped": 0.039713541666666664, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.7984729409217834, + "0.6": 0.7964463233947754, + "0.7": 0.7957195043563843, + "0.8": 0.7957183122634888, + "0.9": 0.7954458594322205 + }, + "unique_frac_signed": { + "0.5": 0.95654296875, + "0.6": 0.98828125, + "0.7": 0.9959309895833334, + "0.8": 0.9977213541666666, + "0.9": 0.99853515625 + }, + "dedup_dcs_signed": { + "0.5": 0.7637736773758661, + "0.6": 0.7871129680424929, + "0.7": 0.7924817134044133, + "0.8": 0.7939051520467425, + "0.9": 0.7942806555365678 + }, + "matched_frac_signed": { + "0.5": 0.7737630208333334, + "0.6": 0.7815755208333334, + "0.7": 0.7584635416666666, + "0.8": 0.7307942708333334, + "0.9": 0.6692708333333334 + }, + "matched_cos_abs": { + "0.5": 0.7997010946273804, + "0.6": 0.7970426678657532, + "0.7": 0.7961682677268982, + "0.8": 0.7963417768478394, + "0.9": 0.7962494492530823 + }, + "unique_frac_abs": { + "0.5": 0.9554036458333334, + "0.6": 0.9874674479166666, + "0.7": 0.99560546875, + "0.8": 0.99755859375, + "0.9": 0.99853515625 + }, + "dedup_dcs_abs": { + "0.5": 0.7640373413839068, + "0.6": 0.7870536891180866, + "0.7": 0.792669481394114, + "0.8": 0.7943975830567069, + "0.9": 0.795083068223903 + }, + "matched_frac_abs": { + "0.5": 0.7724609375, + "0.6": 0.7809244791666666, + "0.7": 0.7604166666666666, + "0.8": 0.7314453125, + "0.9": 0.669921875 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.7987063527107239, + "dcs_abs": 0.7993343472480774, + "frac_sign_flipped": 0.035481770833333336, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.7992627024650574, + "0.6": 0.799367368221283, + "0.7": 0.7988073825836182, + "0.8": 0.7990401387214661, + "0.9": 0.7988802194595337 + }, + "unique_frac_signed": { + "0.5": 0.9542643229166666, + "0.6": 0.9873046875, + "0.7": 0.9954427083333334, + "0.8": 0.9977213541666666, + "0.9": 0.9983723958333334 + }, + "dedup_dcs_signed": { + "0.5": 0.7627078816003632, + "0.6": 0.7892191496794112, + "0.7": 0.7951669843556981, + "0.8": 0.7972194092387023, + "0.9": 0.7975799586856738 + }, + "matched_frac_signed": { + "0.5": 0.7731119791666666, + "0.6": 0.7796223958333334, + "0.7": 0.763671875, + "0.8": 0.7291666666666666, + "0.9": 0.6722005208333334 + }, + "matched_cos_abs": { + "0.5": 0.8003445863723755, + "0.6": 0.799915075302124, + "0.7": 0.7995107173919678, + "0.8": 0.7995234131813049, + "0.9": 0.7995020747184753 + }, + "unique_frac_abs": { + "0.5": 0.953125, + "0.6": 0.9866536458333334, + "0.7": 0.9952799479166666, + "0.8": 0.99755859375, + "0.9": 0.9983723958333334 + }, + "dedup_dcs_abs": { + "0.5": 0.7628284338861704, + "0.6": 0.7892391254038861, + "0.7": 0.7957369851646945, + "0.8": 0.7975714517233428, + "0.9": 0.798200801810405 + }, + "matched_frac_abs": { + "0.5": 0.7727864583333334, + "0.6": 0.7789713541666666, + "0.7": 0.7633463541666666, + "0.8": 0.7294921875, + "0.9": 0.6722005208333334 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 5, + "n_dense_std": 1.5275252316519465, + "l0_mean": 29, + "l0_std": 0.0, + "recon_r2_mean": 0.9341666666666667, + "recon_r2_std": 0.0017616280348965066, + "l1_per_token_mean": 43.4453, + "l1_per_token_std": 0.04491625540937109, + "babel_mu1_signed_mean": 0.35293179750442505, + "babel_mu1_signed_std": 0.0016210962398180726, + "n_dup_signed_per_threshold": { + "0.5": 255.33333333333334, + "0.6": 66.66666666666667, + "0.7": 20.666666666666668, + "0.8": 10.0, + "0.9": 7.0 + }, + "n_dup_signed_band": 26.083333333333336, + "n_dup_signed_band_std": 2.6496855159307744, + "n_dup_clusters_signed_per_threshold": { + "0.5": 120.33333333333333, + "0.6": 29.333333333333332, + "0.7": 7.666666666666667, + "0.8": 3.6666666666666665, + "0.9": 2.6666666666666665 + }, + "n_dup_clusters_signed_band": 10.833333333333332, + "n_dup_clusters_signed_band_std": 0.8779711460710615, + "largest_dup_signed_per_threshold": { + "0.5": 7.0, + "0.6": 6.333333333333333, + "0.7": 6.0, + "0.8": 4.666666666666667, + "0.9": 3.6666666666666665 + }, + "largest_dup_signed_band": 5.166666666666667, + "largest_dup_signed_band_std": 0.8036375634160796, + "dcs_signed_mean": 0.794789989789327, + "dcs_signed_std": 0.004213285292909499, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.7616222127107903, + "0.6": 0.7861045905738138, + "0.7": 0.7918270964033178, + "0.8": 0.7934012452412086, + "0.9": 0.7938071282955611 + }, + "dedup_dcs_signed_band": 0.7912850151284753, + "dedup_dcs_signed_band_std": 0.0038837181948187054, + "matched_cos_signed_per_threshold": { + "0.5": 0.7966337402661642, + "0.6": 0.7957761883735657, + "0.7": 0.7951927582422892, + "0.8": 0.7950413425763448, + "0.9": 0.7949291467666626 + }, + "matched_cos_signed_band": 0.7952348589897156, + "matched_cos_signed_band_std": 0.004120374810524671, + "unique_frac_signed_per_threshold": { + "0.5": 0.9560546875, + "0.6": 0.9878472222222222, + "0.7": 0.9957682291666666, + "0.8": 0.9979383680555555, + "0.9": 0.9985894097222223 + }, + "unique_frac_signed_band": 0.9950358072916666, + "unique_frac_signed_band_std": 0.0002934205139537854, + "matched_frac_signed_per_threshold": { + "0.5": 0.7701822916666666, + "0.6": 0.7768012152777778, + "0.7": 0.7588975694444443, + "0.8": 0.7290581597222222, + "0.9": 0.6698133680555557 + }, + "matched_frac_signed_band": 0.733642578125, + "matched_frac_signed_band_std": 0.0034305401869072537, + "babel_mu1_abs_mean": 0.3540644745031993, + "babel_mu1_abs_std": 0.0015845598584913897, + "n_dup_abs_per_threshold": { + "0.5": 259.3333333333333, + "0.6": 69.66666666666667, + "0.7": 22.333333333333332, + "0.8": 11.333333333333334, + "0.9": 7.0 + }, + "n_dup_abs_band": 27.583333333333332, + "n_dup_abs_band_std": 2.155419526062927, + "n_dup_clusters_abs_per_threshold": { + "0.5": 121.33333333333333, + "0.6": 30.0, + "0.7": 8.333333333333334, + "0.8": 4.333333333333333, + "0.9": 2.6666666666666665 + }, + "n_dup_clusters_abs_band": 11.333333333333334, + "n_dup_clusters_abs_band_std": 0.8779711460710615, + "largest_dup_abs_per_threshold": { + "0.5": 8.0, + "0.6": 6.666666666666667, + "0.7": 6.0, + "0.8": 4.666666666666667, + "0.9": 3.6666666666666665 + }, + "largest_dup_abs_band": 5.250000000000001, + "largest_dup_abs_band_std": 0.6614378277661477, + "dcs_abs_mean": 0.7955158154169718, + "dcs_abs_std": 0.0041535242722263825, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.7618729223872328, + "0.6": 0.7860795535574047, + "0.7": 0.7921746789264338, + "0.8": 0.7938341161643621, + "0.9": 0.79453576876161 + }, + "dedup_dcs_abs_band": 0.7916560293524527, + "dedup_dcs_abs_band_std": 0.0038937899310474687, + "matched_cos_abs_per_threshold": { + "0.5": 0.7977116306622823, + "0.6": 0.7963629563649496, + "0.7": 0.795801560084025, + "0.8": 0.7956478595733643, + "0.9": 0.795658806959788 + }, + "matched_cos_abs_band": 0.7958677957455317, + "matched_cos_abs_band_std": 0.004067824010135765, + "unique_frac_abs_per_threshold": { + "0.5": 0.955078125, + "0.6": 0.9870876736111112, + "0.7": 0.9954427083333334, + "0.8": 0.9977213541666666, + "0.9": 0.9985894097222223 + }, + "unique_frac_abs_band": 0.9947102864583334, + "unique_frac_abs_band_std": 0.00021531179289264017, + "matched_frac_abs_per_threshold": { + "0.5": 0.7696397569444445, + "0.6": 0.7762586805555555, + "0.7": 0.7593315972222222, + "0.8": 0.7292751736111112, + "0.9": 0.6700303819444445 + }, + "matched_frac_abs_band": 0.7337239583333334, + "matched_frac_abs_band_std": 0.003668433486192887, + "alive_source": "stats", + "frac_denominator": "d_sae", + "d_sae": 3072 + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/seed0/config.json b/vision/dinov2-b_layer10/relu/x4/l0_32/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..920481756aae71a20da4a01e54beb062205c045a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..5b3eaac03e2d38c407923d9143a750dfe53fc1ee --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2471, + "r2_score": 0.9331, + "l0_per_token_0": 0.0094, + "l0_per_token_0.001": 0.0094, + "l0_per_token_0.01": 0.0093, + "l0_per_token_0.05": 0.0088, + "l0_per_token_0.1": 0.0083, + "l0_per_token_0.5": 0.0056, + "l0_per_token_1.0": 0.0037, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 43.4364 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/seed1/config.json b/vision/dinov2-b_layer10/relu/x4/l0_32/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2b729114fa7a39fd6d4550e7ebd1e771051ed38a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..0d59207f742ab434003fa64c4858810db276ff2b --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2468, + "r2_score": 0.9332, + "l0_per_token_0": 0.0094, + "l0_per_token_0.001": 0.0094, + "l0_per_token_0.01": 0.0093, + "l0_per_token_0.05": 0.0088, + "l0_per_token_0.1": 0.0084, + "l0_per_token_0.5": 0.0056, + "l0_per_token_1.0": 0.0037, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 43.4055 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/seed2/config.json b/vision/dinov2-b_layer10/relu/x4/l0_32/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..d2ac614f301c676dc8383e32f921621c96254746 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..331cdeade7f37b74b37ffc0e30e1c094a69d9e82 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.2468, + "r2_score": 0.9362, + "l0_per_token_0": 0.0094, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 1.0, + "l1_per_token": 43.494 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/code_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x4/l0_64/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..17884363e7354e3b29808413742af4ed4295e87f --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 66.38440337224384, + "recon_r2": 0.9521496524886579, + "l1_per_token": 63.24008662103471 + }, + "seed1": { + "l0": 66.13775097276265, + "recon_r2": 0.9521067327006207, + "l1_per_token": 63.13545321411877 + }, + "seed2": { + "l0": 66.88821530479896, + "recon_r2": 0.952173898946571, + "l1_per_token": 63.38919688796638 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.8468062294234072, + "mcc_abs": 0.8468306157815659, + "act_iou": 0.6150470695634432, + "act_iou_n_pairs": 3072, + "input_r2": 0.9945961419777345, + "matched_mcc_signed": { + "0.6": 0.9444594343238162, + "0.7": 0.9553389086933861, + "0.8": 0.9698430599586132, + "0.9": 0.9820203432477086 + }, + "matched_mcc_abs": { + "0.6": 0.9445633124551577, + "0.7": 0.9555086678930979, + "0.8": 0.9698792601854904, + "0.9": 0.9820045391742178 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.8472440711561448, + "mcc_abs": 0.8472741545564434, + "act_iou": 0.6134884222085363, + "act_iou_n_pairs": 3072, + "input_r2": 0.9945644154013099, + "matched_mcc_signed": { + "0.6": 0.9386993468900303, + "0.7": 0.9529986704046098, + "0.8": 0.9685094443131542, + "0.9": 0.9807003677957427 + }, + "matched_mcc_abs": { + "0.6": 0.9380750180614505, + "0.7": 0.9525992401352407, + "0.8": 0.9684709097406811, + "0.9": 0.9807003677957427 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.8463393051372589, + "mcc_abs": 0.8463538499047911, + "act_iou": 0.6146141636675034, + "act_iou_n_pairs": 3072, + "input_r2": 0.9945305533278209, + "matched_mcc_signed": { + "0.6": 0.9408369752141642, + "0.7": 0.9548828590686856, + "0.8": 0.969316012013835, + "0.9": 0.9808018752249172 + }, + "matched_mcc_abs": { + "0.6": 0.9405786188774156, + "0.7": 0.9548539269304317, + "0.8": 0.9691930276447261, + "0.9": 0.9808095166371293 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 66.4701232166018, + "l0_std": 0.3825050289894831, + "recon_r2_mean": 0.9521434280452832, + "recon_r2_std": 3.4012996343404975e-05, + "l1_per_token_mean": 63.25491224103996, + "l1_per_token_std": 0.12751985045797873, + "input_r2_mean": 0.9945637035689551, + "input_r2_std": 3.2800118572153165e-05, + "mcc_signed_mean": 0.8467965352389369, + "mcc_signed_std": 0.00045246090454542626, + "mcc_abs_mean": 0.8468195400809334, + "mcc_abs_std": 0.00046025228551477775, + "matched_mcc_signed_per_threshold": { + "0.6": 0.9413319188093369, + "0.7": 0.9544068127222273, + "0.8": 0.9692228387618674, + "0.9": 0.9811741954227896 + }, + "matched_mcc_signed_band": 0.9615339414290553, + "matched_mcc_abs_per_threshold": { + "0.6": 0.9410723164646746, + "0.7": 0.9543206116529234, + "0.8": 0.969181065856966, + "0.9": 0.9811714745356966 + }, + "matched_mcc_abs_band": 0.9614363671275652, + "act_iou_mean": 0.6143832184798277, + "act_iou_std": 0.0008045788674968842, + "act_iou_n_pairs": 3072.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/dict_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x4/l0_64/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..330a14137d80ded1429dce06e57c322e77aa9cf9 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/dict_metrics_step_final.json @@ -0,0 +1,545 @@ +{ + "schema": 2, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 3072, + "d_in": 768, + "l0": 66, + "l0_raw": 66.3552, + "recon_r2": 0.9521, + "l1_per_token": 63.2571, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 25, + "dense_rate": 0.008138020833333334, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 117, + "0.6": 14, + "0.7": 9, + "0.8": 6, + "0.9": 4 + }, + "n_dup_clusters_signed": { + "0.5": 56, + "0.6": 5, + "0.7": 3, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_signed": { + "0.5": 6, + "0.6": 5, + "0.7": 5, + "0.8": 4, + "0.9": 4 + }, + "babel_mu1_signed": 0.307264119386673, + "n_dup_abs": { + "0.5": 148, + "0.6": 18, + "0.7": 11, + "0.8": 6, + "0.9": 4 + }, + "n_dup_clusters_abs": { + "0.5": 71, + "0.6": 7, + "0.7": 4, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_abs": { + "0.5": 6, + "0.6": 5, + "0.7": 5, + "0.8": 4, + "0.9": 4 + }, + "babel_mu1_abs": 0.3125115931034088 + }, + "seed1": { + "d_sae": 3072, + "d_in": 768, + "l0": 66, + "l0_raw": 66.048, + "recon_r2": 0.9521, + "l1_per_token": 63.1502, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 18, + "dense_rate": 0.005859375, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 93, + "0.6": 25, + "0.7": 11, + "0.8": 5, + "0.9": 3 + }, + "n_dup_clusters_signed": { + "0.5": 44, + "0.6": 11, + "0.7": 4, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_signed": { + "0.5": 6, + "0.6": 4, + "0.7": 4, + "0.8": 3, + "0.9": 3 + }, + "babel_mu1_signed": 0.3057239353656769, + "n_dup_abs": { + "0.5": 116, + "0.6": 27, + "0.7": 13, + "0.8": 5, + "0.9": 3 + }, + "n_dup_clusters_abs": { + "0.5": 54, + "0.6": 12, + "0.7": 5, + "0.8": 2, + "0.9": 1 + }, + "largest_dup_abs": { + "0.5": 7, + "0.6": 4, + "0.7": 4, + "0.8": 3, + "0.9": 3 + }, + "babel_mu1_abs": 0.3104725778102875 + }, + "seed2": { + "d_sae": 3072, + "d_in": 768, + "l0": 67, + "l0_raw": 66.9696, + "recon_r2": 0.9521, + "l1_per_token": 63.413, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 24, + "dense_rate": 0.0078125, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 107, + "0.6": 16, + "0.7": 10, + "0.8": 5, + "0.9": 5 + }, + "n_dup_clusters_signed": { + "0.5": 50, + "0.6": 6, + "0.7": 3, + "0.8": 2, + "0.9": 2 + }, + "largest_dup_signed": { + "0.5": 7, + "0.6": 5, + "0.7": 5, + "0.8": 3, + "0.9": 3 + }, + "babel_mu1_signed": 0.3071190118789673, + "n_dup_abs": { + "0.5": 128, + "0.6": 23, + "0.7": 12, + "0.8": 7, + "0.9": 5 + }, + "n_dup_clusters_abs": { + "0.5": 59, + "0.6": 9, + "0.7": 4, + "0.8": 3, + "0.9": 2 + }, + "largest_dup_abs": { + "0.5": 8, + "0.6": 5, + "0.7": 5, + "0.8": 3, + "0.9": 3 + }, + "babel_mu1_abs": 0.31253179907798767 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.8759455680847168, + "dcs_abs": 0.8761868476867676, + "frac_sign_flipped": 0.0078125, + "alive_source": "stats", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8788420557975769, + "0.6": 0.8769009113311768, + "0.7": 0.8759840726852417, + "0.8": 0.8760386109352112, + "0.9": 0.8760771155357361 + }, + "unique_frac_signed": { + "0.5": 0.9820963541666666, + "0.6": 0.9962565104166666, + "0.7": 0.9978841145833334, + "0.8": 0.9988606770833334, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_signed": { + "0.5": 0.8631075788871385, + "0.6": 0.873618241903993, + "0.7": 0.8741305907606147, + "0.8": 0.8750405200698879, + "0.9": 0.8753641621539524 + }, + "matched_frac_signed": { + "0.5": 0.8776041666666666, + "0.6": 0.861328125, + "0.7": 0.83984375, + "0.8": 0.802734375, + "0.9": 0.7513020833333334 + }, + "matched_cos_abs": { + "0.5": 0.88028484582901, + "0.6": 0.8768486380577087, + "0.7": 0.8761575818061829, + "0.8": 0.8762308359146118, + "0.9": 0.8762691617012024 + }, + "unique_frac_abs": { + "0.5": 0.9773763020833334, + "0.6": 0.9957682291666666, + "0.7": 0.99755859375, + "0.8": 0.9988606770833334, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_abs": { + "0.5": 0.8603695473963551, + "0.6": 0.8731380155659281, + "0.7": 0.8740185252099764, + "0.8": 0.8752325260429643, + "0.9": 0.8755560520318492 + }, + "matched_frac_abs": { + "0.5": 0.875, + "0.6": 0.8603515625, + "0.7": 0.8391927083333334, + "0.8": 0.802734375, + "0.9": 0.7513020833333334 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.8767868876457214, + "dcs_abs": 0.8771607279777527, + "frac_sign_flipped": 0.0078125, + "alive_source": "stats", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8781667947769165, + "0.6": 0.8766931891441345, + "0.7": 0.8766819834709167, + "0.8": 0.876816987991333, + "0.9": 0.8766517043113708 + }, + "unique_frac_signed": { + "0.5": 0.9807942708333334, + "0.6": 0.9969075520833334, + "0.7": 0.9978841145833334, + "0.8": 0.9988606770833334, + "0.9": 0.9990234375 + }, + "dedup_dcs_signed": { + "0.5": 0.8613009611532714, + "0.6": 0.87398206111781, + "0.7": 0.8748270248470362, + "0.8": 0.8758180103031918, + "0.9": 0.8757955991313793 + }, + "matched_frac_signed": { + "0.5": 0.8834635416666666, + "0.6": 0.8681640625, + "0.7": 0.8411458333333334, + "0.8": 0.8040364583333334, + "0.9": 0.74609375 + }, + "matched_cos_abs": { + "0.5": 0.8797926306724548, + "0.6": 0.8773223161697388, + "0.7": 0.877050518989563, + "0.8": 0.8770760893821716, + "0.9": 0.8770259022712708 + }, + "unique_frac_abs": { + "0.5": 0.9762369791666666, + "0.6": 0.9959309895833334, + "0.7": 0.99755859375, + "0.8": 0.9986979166666666, + "0.9": 0.9990234375 + }, + "dedup_dcs_abs": { + "0.5": 0.8588861000607722, + "0.6": 0.87375248252647, + "0.7": 0.8749092823709361, + "0.8": 0.8759340632241219, + "0.9": 0.876169431663584 + }, + "matched_frac_abs": { + "0.5": 0.8792317708333334, + "0.6": 0.8678385416666666, + "0.7": 0.8411458333333334, + "0.8": 0.8037109375, + "0.9": 0.74609375 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.8756335377693176, + "dcs_abs": 0.8758605122566223, + "frac_sign_flipped": 0.0068359375, + "alive_source": "stats", + "n_alive_a": 3072, + "n_alive_b": 3072, + "matched_cos_signed": { + "0.5": 0.8777368664741516, + "0.6": 0.8761578798294067, + "0.7": 0.875525712966919, + "0.8": 0.8755146861076355, + "0.9": 0.8756418228149414 + }, + "unique_frac_signed": { + "0.5": 0.9827473958333334, + "0.6": 0.99609375, + "0.7": 0.9977213541666666, + "0.8": 0.9990234375, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_signed": { + "0.5": 0.8625936197543828, + "0.6": 0.8727353881113231, + "0.7": 0.8735306999490906, + "0.8": 0.8746596912969835, + "0.9": 0.8749292236752808 + }, + "matched_frac_signed": { + "0.5": 0.8854166666666666, + "0.6": 0.861328125, + "0.7": 0.8349609375, + "0.8": 0.8004557291666666, + "0.9": 0.7457682291666666 + }, + "matched_cos_abs": { + "0.5": 0.8785207271575928, + "0.6": 0.8761000037193298, + "0.7": 0.8757866024971008, + "0.8": 0.8759366273880005, + "0.9": 0.8758885860443115 + }, + "unique_frac_abs": { + "0.5": 0.9786783854166666, + "0.6": 0.9952799479166666, + "0.7": 0.9973958333333334, + "0.8": 0.9988606770833334, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_abs": { + "0.5": 0.8597892468096688, + "0.6": 0.871964766071566, + "0.7": 0.8735059082197647, + "0.8": 0.8749386527148697, + "0.9": 0.8751757860882208 + }, + "matched_frac_abs": { + "0.5": 0.8828125, + "0.6": 0.8610026041666666, + "0.7": 0.8336588541666666, + "0.8": 0.7998046875, + "0.9": 0.7454427083333334 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 22, + "n_dense_std": 3.7859388972001824, + "l0_mean": 66, + "l0_std": 0.4692557511634775, + "recon_r2_mean": 0.9521000000000001, + "recon_r2_std": 5.773502691895622e-05, + "l1_per_token_mean": 63.27343333333332, + "l1_per_token_std": 0.13184854948007582, + "babel_mu1_signed_mean": 0.3067023555437724, + "babel_mu1_signed_std": 0.0008504372880327279, + "n_dup_signed_per_threshold": { + "0.5": 105.66666666666667, + "0.6": 18.333333333333332, + "0.7": 10.0, + "0.8": 5.333333333333333, + "0.9": 4.0 + }, + "n_dup_signed_band": 9.416666666666666, + "n_dup_signed_band_std": 1.4215601757693317, + "n_dup_clusters_signed_per_threshold": { + "0.5": 50.0, + "0.6": 7.333333333333333, + "0.7": 3.3333333333333335, + "0.8": 2.0, + "0.9": 1.3333333333333333 + }, + "n_dup_clusters_signed_band": 3.5, + "n_dup_clusters_signed_band_std": 0.9013878188659973, + "largest_dup_signed_per_threshold": { + "0.5": 6.333333333333333, + "0.6": 4.666666666666667, + "0.7": 4.666666666666667, + "0.8": 3.3333333333333335, + "0.9": 3.3333333333333335 + }, + "largest_dup_signed_band": 4.0, + "largest_dup_signed_band_std": 0.5, + "dcs_signed_mean": 0.876121997833252, + "dcs_signed_std": 0.0005965731945673591, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.8623340532649308, + "0.6": 0.8734452303777087, + "0.7": 0.8741627718522471, + "0.8": 0.8751727405566877, + "0.9": 0.8753629949868708 + }, + "dedup_dcs_signed_band": 0.8745359344433785, + "dedup_dcs_signed_band_std": 0.0005709654698006789, + "matched_cos_signed_per_threshold": { + "0.5": 0.8782485723495483, + "0.6": 0.876583993434906, + "0.7": 0.8760639230410258, + "0.8": 0.8761234283447266, + "0.9": 0.8761235475540161 + }, + "matched_cos_signed_band": 0.8762237230936686, + "matched_cos_signed_band_std": 0.0005009945136208848, + "unique_frac_signed_per_threshold": { + "0.5": 0.9818793402777778, + "0.6": 0.9964192708333334, + "0.7": 0.9978298611111112, + "0.8": 0.9989149305555557, + "0.9": 0.9991319444444443 + }, + "unique_frac_signed_band": 0.9980740017361112, + "unique_frac_signed_band_std": 8.470320635851485e-05, + "matched_frac_signed_per_threshold": { + "0.5": 0.8821614583333334, + "0.6": 0.8636067708333334, + "0.7": 0.8386501736111112, + "0.8": 0.8024088541666666, + "0.9": 0.7477213541666666 + }, + "matched_frac_signed_band": 0.8130967881944444, + "matched_frac_signed_band_std": 0.0022022833653206927, + "babel_mu1_abs_mean": 0.31183865666389465, + "babel_mu1_abs_std": 0.0011831021283290365, + "n_dup_abs_per_threshold": { + "0.5": 130.66666666666666, + "0.6": 22.666666666666668, + "0.7": 12.0, + "0.8": 6.0, + "0.9": 4.0 + }, + "n_dup_abs_band": 11.166666666666668, + "n_dup_abs_band_std": 1.233220715579062, + "n_dup_clusters_abs_per_threshold": { + "0.5": 61.333333333333336, + "0.6": 9.333333333333334, + "0.7": 4.333333333333333, + "0.8": 2.3333333333333335, + "0.9": 1.3333333333333333 + }, + "n_dup_clusters_abs_band": 4.333333333333333, + "n_dup_clusters_abs_band_std": 0.7637626158259734, + "largest_dup_abs_per_threshold": { + "0.5": 7.0, + "0.6": 4.666666666666667, + "0.7": 4.666666666666667, + "0.8": 3.3333333333333335, + "0.9": 3.3333333333333335 + }, + "largest_dup_abs_band": 4.0, + "largest_dup_abs_band_std": 0.5, + "dcs_abs_mean": 0.8764026959737142, + "dcs_abs_std": 0.0006764488839535355, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.8596816314222654, + "0.6": 0.8729517547213214, + "0.7": 0.8741445719335591, + "0.8": 0.8753684139939854, + "0.9": 0.8756337565945512 + }, + "dedup_dcs_abs_band": 0.8745246243108542, + "dedup_dcs_abs_band_std": 0.0006483692829311848, + "matched_cos_abs_per_threshold": { + "0.5": 0.8795327345530192, + "0.6": 0.8767569859822592, + "0.7": 0.8763315677642822, + "0.8": 0.8764145175615946, + "0.9": 0.8763945500055949 + }, + "matched_cos_abs_band": 0.8764744053284327, + "matched_cos_abs_band_std": 0.0006013763917074658, + "unique_frac_abs_per_threshold": { + "0.5": 0.9774305555555555, + "0.6": 0.9956597222222222, + "0.7": 0.9975043402777778, + "0.8": 0.9988064236111112, + "0.9": 0.9991319444444443 + }, + "unique_frac_abs_band": 0.9977756076388888, + "unique_frac_abs_band_std": 8.47032063584704e-05, + "matched_frac_abs_per_threshold": { + "0.5": 0.8790147569444445, + "0.6": 0.863064236111111, + "0.7": 0.8379991319444445, + "0.8": 0.8020833333333334, + "0.9": 0.7476128472222223 + }, + "matched_frac_abs_band": 0.8126898871527778, + "matched_frac_abs_band_std": 0.0024377866686915426, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/seed0/config.json b/vision/dinov2-b_layer10/relu/x4/l0_64/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..920481756aae71a20da4a01e54beb062205c045a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..7411c0bcfab24ae58020ff24949644dc7a2e4429 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1768, + "r2_score": 0.9521, + "l0_per_token_0": 0.0216, + "l0_per_token_0.001": 0.0216, + "l0_per_token_0.01": 0.0212, + "l0_per_token_0.05": 0.0197, + "l0_per_token_0.1": 0.0181, + "l0_per_token_0.5": 0.0098, + "l0_per_token_1.0": 0.0054, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0003, + "live_latent_frac": 1.0, + "l1_per_token": 63.2571 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/seed1/config.json b/vision/dinov2-b_layer10/relu/x4/l0_64/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2b729114fa7a39fd6d4550e7ebd1e771051ed38a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..dc1a15742431b1bcdd708c20f4cd9c36ef1fbdb6 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1769, + "r2_score": 0.9521, + "l0_per_token_0": 0.0215, + "l0_per_token_0.001": 0.0215, + "l0_per_token_0.01": 0.0212, + "l0_per_token_0.05": 0.0197, + "l0_per_token_0.1": 0.018, + "l0_per_token_0.5": 0.0097, + "l0_per_token_1.0": 0.0054, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0007, + "live_latent_frac": 1.0, + "l1_per_token": 63.1502 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/seed2/config.json b/vision/dinov2-b_layer10/relu/x4/l0_64/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..d2ac614f301c676dc8383e32f921621c96254746 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 4, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..530a89bca7eaa2828fb15215f5db1127d53650d1 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x4/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1767, + "r2_score": 0.9521, + "l0_per_token_0": 0.0218, + "l0_per_token_0.001": 0.0217, + "l0_per_token_0.01": 0.0214, + "l0_per_token_0.05": 0.0199, + "l0_per_token_0.1": 0.0182, + "l0_per_token_0.5": 0.0098, + "l0_per_token_1.0": 0.0054, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0003, + "live_latent_frac": 1.0, + "l1_per_token": 63.413 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/code_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x8/l0_128/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..db9302c62cb9084c2a293c2f8d0b56e2bc2dd81b --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 127.58820492866407, + "recon_r2": 0.9643444611789529, + "l1_per_token": 74.57088104464506 + }, + "seed1": { + "l0": 126.8789662775616, + "recon_r2": 0.9642873710930381, + "l1_per_token": 74.33392074982693 + }, + "seed2": { + "l0": 127.63566796368353, + "recon_r2": 0.9645063000221357, + "l1_per_token": 74.60909932508298 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.6860603589116817, + "mcc_abs": 0.6860628665423519, + "act_iou": 0.477440112095229, + "act_iou_n_pairs": 6144, + "input_r2": 0.9948605841609822, + "matched_mcc_signed": { + "0.6": 0.9015153889750307, + "0.7": 0.928998693912241, + "0.8": 0.9559000390364512, + "0.9": 0.977262924825148 + }, + "matched_mcc_abs": { + "0.6": 0.901772707479833, + "0.7": 0.9292000922787992, + "0.8": 0.9561546674660926, + "0.9": 0.9771171858715338 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6855920037372623, + "mcc_abs": 0.6855967081445634, + "act_iou": 0.4723063497433942, + "act_iou_n_pairs": 6144, + "input_r2": 0.994668268438609, + "matched_mcc_signed": { + "0.6": 0.9006491438480261, + "0.7": 0.9281623355020721, + "0.8": 0.9561555233239256, + "0.9": 0.9775179144457414 + }, + "matched_mcc_abs": { + "0.6": 0.9009868527427827, + "0.7": 0.9285556674073463, + "0.8": 0.9561699636771316, + "0.9": 0.9774555514945327 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6832130414311575, + "mcc_abs": 0.6832179498976565, + "act_iou": 0.46896443623386525, + "act_iou_n_pairs": 6144, + "input_r2": 0.994586705565264, + "matched_mcc_signed": { + "0.6": 0.9045903380568774, + "0.7": 0.9326064034591333, + "0.8": 0.9581489775296615, + "0.9": 0.9776708213713894 + }, + "matched_mcc_abs": { + "0.6": 0.9046922642994288, + "0.7": 0.9323111509387124, + "0.8": 0.958044680754327, + "0.9": 0.9774207075463942 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 127.3676130566364, + "l0_std": 0.423845421054004, + "recon_r2_mean": 0.9643793774313756, + "recon_r2_std": 0.00011356419995485163, + "l1_per_token_mean": 74.50463370651833, + "l1_per_token_std": 0.14907160838552097, + "input_r2_mean": 0.9947051860549517, + "input_r2_std": 0.0001406220256413363, + "mcc_signed_mean": 0.6849551346933671, + "mcc_signed_std": 0.00152676319153582, + "mcc_abs_mean": 0.6849591748615239, + "mcc_abs_std": 0.0015258519567944984, + "matched_mcc_signed_per_threshold": { + "0.6": 0.9022516236266447, + "0.7": 0.9299224776244821, + "0.8": 0.9567348466300127, + "0.9": 0.9774838868807597 + }, + "matched_mcc_signed_band": 0.9415982086904748, + "matched_mcc_abs_per_threshold": { + "0.6": 0.9024839415073481, + "0.7": 0.9300223035416194, + "0.8": 0.9567897706325171, + "0.9": 0.9773311483041537 + }, + "matched_mcc_abs_band": 0.9416567909964096, + "act_iou_mean": 0.4729036326908295, + "act_iou_std": 0.0042692892284546475, + "act_iou_n_pairs": 6144.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/dict_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x8/l0_128/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..57b344fca2bdfe3aecfa38b6858ec21096bba5bc --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/dict_metrics_step_final.json @@ -0,0 +1,545 @@ +{ + "schema": 2, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 6144, + "d_in": 768, + "l0": 128, + "l0_raw": 127.7952, + "recon_r2": 0.9643, + "l1_per_token": 74.5924, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 9, + "dense_rate": 0.00146484375, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 466, + "0.6": 132, + "0.7": 65, + "0.8": 27, + "0.9": 7 + }, + "n_dup_clusters_signed": { + "0.5": 225, + "0.6": 62, + "0.7": 30, + "0.8": 12, + "0.9": 3 + }, + "largest_dup_signed": { + "0.5": 8, + "0.6": 7, + "0.7": 7, + "0.8": 5, + "0.9": 3 + }, + "babel_mu1_signed": 0.3470424711704254, + "n_dup_abs": { + "0.5": 479, + "0.6": 140, + "0.7": 67, + "0.8": 29, + "0.9": 7 + }, + "n_dup_clusters_abs": { + "0.5": 231, + "0.6": 65, + "0.7": 31, + "0.8": 13, + "0.9": 3 + }, + "largest_dup_abs": { + "0.5": 8, + "0.6": 7, + "0.7": 6, + "0.8": 5, + "0.9": 3 + }, + "babel_mu1_abs": 0.3550845682621002 + }, + "seed1": { + "d_sae": 6144, + "d_in": 768, + "l0": 127, + "l0_raw": 127.1808, + "recon_r2": 0.9643, + "l1_per_token": 74.3676, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 7, + "dense_rate": 0.0011393229166666667, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 528, + "0.6": 157, + "0.7": 54, + "0.8": 22, + "0.9": 14 + }, + "n_dup_clusters_signed": { + "0.5": 256, + "0.6": 74, + "0.7": 25, + "0.8": 10, + "0.9": 6 + }, + "largest_dup_signed": { + "0.5": 7, + "0.6": 6, + "0.7": 6, + "0.8": 4, + "0.9": 4 + }, + "babel_mu1_signed": 0.3482103645801544, + "n_dup_abs": { + "0.5": 545, + "0.6": 158, + "0.7": 56, + "0.8": 24, + "0.9": 16 + }, + "n_dup_clusters_abs": { + "0.5": 264, + "0.6": 74, + "0.7": 26, + "0.8": 11, + "0.9": 7 + }, + "largest_dup_abs": { + "0.5": 7, + "0.6": 7, + "0.7": 6, + "0.8": 4, + "0.9": 4 + }, + "babel_mu1_abs": 0.3561321496963501 + }, + "seed2": { + "d_sae": 6144, + "d_in": 768, + "l0": 128, + "l0_raw": 127.7952, + "recon_r2": 0.9645, + "l1_per_token": 74.616, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 8, + "dense_rate": 0.0013020833333333333, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 457, + "0.6": 130, + "0.7": 47, + "0.8": 23, + "0.9": 8 + }, + "n_dup_clusters_signed": { + "0.5": 222, + "0.6": 62, + "0.7": 21, + "0.8": 10, + "0.9": 3 + }, + "largest_dup_signed": { + "0.5": 6, + "0.6": 6, + "0.7": 6, + "0.8": 5, + "0.9": 4 + }, + "babel_mu1_signed": 0.34473729133605957, + "n_dup_abs": { + "0.5": 471, + "0.6": 131, + "0.7": 49, + "0.8": 25, + "0.9": 10 + }, + "n_dup_clusters_abs": { + "0.5": 229, + "0.6": 62, + "0.7": 22, + "0.8": 11, + "0.9": 4 + }, + "largest_dup_abs": { + "0.5": 7, + "0.6": 6, + "0.7": 6, + "0.8": 5, + "0.9": 4 + }, + "babel_mu1_abs": 0.3531961739063263 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.7572271227836609, + "dcs_abs": 0.7576702237129211, + "frac_sign_flipped": 0.017252604166666668, + "alive_source": "stats", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.7604324221611023, + "0.6": 0.7602892518043518, + "0.7": 0.7590299844741821, + "0.8": 0.7580401301383972, + "0.9": 0.757452666759491 + }, + "unique_frac_signed": { + "0.5": 0.958251953125, + "0.6": 0.987548828125, + "0.7": 0.9947916666666666, + "0.8": 0.997802734375, + "0.9": 0.9990234375 + }, + "dedup_dcs_signed": { + "0.5": 0.7286858537554508, + "0.6": 0.7508227596554207, + "0.7": 0.7550767033050457, + "0.8": 0.7563745146180736, + "0.9": 0.7567129668896087 + }, + "matched_frac_signed": { + "0.5": 0.7088216145833334, + "0.6": 0.6881510416666666, + "0.7": 0.6437174479166666, + "0.8": 0.58935546875, + "0.9": 0.5216471354166666 + }, + "matched_cos_abs": { + "0.5": 0.7608625888824463, + "0.6": 0.7603538036346436, + "0.7": 0.7594289183616638, + "0.8": 0.7584547400474548, + "0.9": 0.7579395174980164 + }, + "unique_frac_abs": { + "0.5": 0.9569498697916666, + "0.6": 0.987060546875, + "0.7": 0.99462890625, + "0.8": 0.9976399739583334, + "0.9": 0.9989420572916666 + }, + "dedup_dcs_abs": { + "0.5": 0.7281073553604074, + "0.6": 0.7505152412340976, + "0.7": 0.7553499544446822, + "0.8": 0.7566647671095174, + "0.9": 0.7571376609121216 + }, + "matched_frac_abs": { + "0.5": 0.70703125, + "0.6": 0.6871744791666666, + "0.7": 0.6433919270833334, + "0.8": 0.5888671875, + "0.9": 0.5218098958333334 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.7550984025001526, + "dcs_abs": 0.7555964589118958, + "frac_sign_flipped": 0.0205078125, + "alive_source": "stats", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.7564693093299866, + "0.6": 0.7573115229606628, + "0.7": 0.7567442059516907, + "0.8": 0.7556754350662231, + "0.9": 0.7552014589309692 + }, + "unique_frac_signed": { + "0.5": 0.9612630208333334, + "0.6": 0.98876953125, + "0.7": 0.9950358072916666, + "0.8": 0.9977213541666666, + "0.9": 0.999267578125 + }, + "dedup_dcs_signed": { + "0.5": 0.7271659734542482, + "0.6": 0.7488065595680382, + "0.7": 0.7529875818824318, + "0.8": 0.7539535183847571, + "0.9": 0.7546483328624163 + }, + "matched_frac_signed": { + "0.5": 0.71142578125, + "0.6": 0.6891276041666666, + "0.7": 0.64453125, + "0.8": 0.5870768229166666, + "0.9": 0.5183919270833334 + }, + "matched_cos_abs": { + "0.5": 0.7568533420562744, + "0.6": 0.757956862449646, + "0.7": 0.7572228312492371, + "0.8": 0.7561330199241638, + "0.9": 0.7557259798049927 + }, + "unique_frac_abs": { + "0.5": 0.9601236979166666, + "0.6": 0.98828125, + "0.7": 0.994873046875, + "0.8": 0.99755859375, + "0.9": 0.9991861979166666 + }, + "dedup_dcs_abs": { + "0.5": 0.7266728295556579, + "0.6": 0.7490745554678142, + "0.7": 0.7533405852882424, + "0.8": 0.7542869920434896, + "0.9": 0.7551109684281982 + }, + "matched_frac_abs": { + "0.5": 0.7111002604166666, + "0.6": 0.6886393229166666, + "0.7": 0.6442057291666666, + "0.8": 0.5867513020833334, + "0.9": 0.5182291666666666 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.7526457905769348, + "dcs_abs": 0.7531244158744812, + "frac_sign_flipped": 0.020345052083333332, + "alive_source": "stats", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.7555974125862122, + "0.6": 0.7552086710929871, + "0.7": 0.753685712814331, + "0.8": 0.7528784275054932, + "0.9": 0.7528560161590576 + }, + "unique_frac_signed": { + "0.5": 0.958740234375, + "0.6": 0.9877115885416666, + "0.7": 0.9955240885416666, + "0.8": 0.9979654947916666, + "0.9": 0.9989420572916666 + }, + "dedup_dcs_signed": { + "0.5": 0.7244216404360486, + "0.6": 0.7459283562056953, + "0.7": 0.7503122822963633, + "0.8": 0.7513466924234914, + "0.9": 0.7520595376263373 + }, + "matched_frac_signed": { + "0.5": 0.7071940104166666, + "0.6": 0.6778971354166666, + "0.7": 0.6349283854166666, + "0.8": 0.5818684895833334, + "0.9": 0.5193684895833334 + }, + "matched_cos_abs": { + "0.5": 0.7561522722244263, + "0.6": 0.7555886507034302, + "0.7": 0.7541199922561646, + "0.8": 0.7533242702484131, + "0.9": 0.7532984018325806 + }, + "unique_frac_abs": { + "0.5": 0.9574381510416666, + "0.6": 0.987548828125, + "0.7": 0.995361328125, + "0.8": 0.997802734375, + "0.9": 0.998779296875 + }, + "dedup_dcs_abs": { + "0.5": 0.7239690334245097, + "0.6": 0.7461806865467224, + "0.7": 0.7506218770577107, + "0.8": 0.751669016724918, + "0.9": 0.752378848119406 + }, + "matched_frac_abs": { + "0.5": 0.70654296875, + "0.6": 0.677734375, + "0.7": 0.6355794270833334, + "0.8": 0.5818684895833334, + "0.9": 0.5196940104166666 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 8, + "n_dense_std": 1.0, + "l0_mean": 128, + "l0_std": 0.35472400539009985, + "recon_r2_mean": 0.9643666666666667, + "recon_r2_std": 0.00011547005383791244, + "l1_per_token_mean": 74.52533333333332, + "l1_per_token_std": 0.1365748512721129, + "babel_mu1_signed_mean": 0.34666337569554645, + "babel_mu1_signed_std": 0.0017672986374739131, + "n_dup_signed_per_threshold": { + "0.5": 483.6666666666667, + "0.6": 139.66666666666666, + "0.7": 55.333333333333336, + "0.8": 24.0, + "0.9": 9.666666666666666 + }, + "n_dup_signed_band": 57.166666666666664, + "n_dup_signed_band_std": 4.901105317510871, + "n_dup_clusters_signed_per_threshold": { + "0.5": 234.33333333333334, + "0.6": 66.0, + "0.7": 25.333333333333332, + "0.8": 10.666666666666666, + "0.9": 4.0 + }, + "n_dup_clusters_signed_band": 26.5, + "n_dup_clusters_signed_band_std": 2.384848003542364, + "largest_dup_signed_per_threshold": { + "0.5": 7.0, + "0.6": 6.333333333333333, + "0.7": 6.333333333333333, + "0.8": 4.666666666666667, + "0.9": 3.6666666666666665 + }, + "largest_dup_signed_band": 5.25, + "largest_dup_signed_band_std": 0.25, + "dcs_signed_mean": 0.7549904386202494, + "dcs_signed_std": 0.002292573520439248, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.7267578225485826, + "0.6": 0.7485192251430514, + "0.7": 0.7527921891612802, + "0.8": 0.7538915751421073, + "0.9": 0.7544736124594541 + }, + "dedup_dcs_signed_band": 0.7524191504764732, + "dedup_dcs_signed_band_std": 0.00242252162587264, + "matched_cos_signed_per_threshold": { + "0.5": 0.7574997146924337, + "0.6": 0.7576031486193339, + "0.7": 0.7564866344134012, + "0.8": 0.7555313309033712, + "0.9": 0.7551700472831726 + }, + "matched_cos_signed_band": 0.7561977903048197, + "matched_cos_signed_band_std": 0.002523086597754289, + "unique_frac_signed_per_threshold": { + "0.5": 0.9594184027777778, + "0.6": 0.9880099826388888, + "0.7": 0.9951171875, + "0.8": 0.997829861111111, + "0.9": 0.9990776909722222 + }, + "unique_frac_signed_band": 0.9950086805555556, + "unique_frac_signed_band_std": 0.00020480236640815205, + "matched_frac_signed_per_threshold": { + "0.5": 0.7091471354166666, + "0.6": 0.68505859375, + "0.7": 0.6410590277777777, + "0.8": 0.5861002604166666, + "0.9": 0.5198025173611112 + }, + "matched_frac_signed_band": 0.6080050998263888, + "matched_frac_signed_band_std": 0.003916057002637385, + "babel_mu1_abs_mean": 0.35480429728825885, + "babel_mu1_abs_std": 0.0014879187894454987, + "n_dup_abs_per_threshold": { + "0.5": 498.3333333333333, + "0.6": 143.0, + "0.7": 57.333333333333336, + "0.8": 26.0, + "0.9": 11.0 + }, + "n_dup_abs_band": 59.333333333333336, + "n_dup_abs_band_std": 5.027010377285224, + "n_dup_clusters_abs_per_threshold": { + "0.5": 241.33333333333334, + "0.6": 67.0, + "0.7": 26.333333333333332, + "0.8": 11.666666666666666, + "0.9": 4.666666666666667 + }, + "n_dup_clusters_abs_band": 27.416666666666668, + "n_dup_clusters_abs_band_std": 2.4281337140555777, + "largest_dup_abs_per_threshold": { + "0.5": 7.333333333333333, + "0.6": 6.666666666666667, + "0.7": 6.0, + "0.8": 4.666666666666667, + "0.9": 3.6666666666666665 + }, + "largest_dup_abs_band": 5.250000000000001, + "largest_dup_abs_band_std": 0.0, + "dcs_abs_mean": 0.755463699499766, + "dcs_abs_std": 0.0022758099705681217, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.7262497394468582, + "0.6": 0.7485901610828781, + "0.7": 0.7531041389302118, + "0.8": 0.7542069252926417, + "0.9": 0.7548758258199086 + }, + "dedup_dcs_abs_band": 0.75269426278141, + "dedup_dcs_abs_band_std": 0.002362820866376455, + "matched_cos_abs_per_threshold": { + "0.5": 0.757956067721049, + "0.6": 0.7579664389292399, + "0.7": 0.7569239139556885, + "0.8": 0.7559706767400106, + "0.9": 0.7556546330451965 + }, + "matched_cos_abs_band": 0.7566289156675339, + "matched_cos_abs_band_std": 0.002483291298279052, + "unique_frac_abs_per_threshold": { + "0.5": 0.9581705729166666, + "0.6": 0.9876302083333334, + "0.7": 0.9949544270833334, + "0.8": 0.9976671006944445, + "0.9": 0.9989691840277777 + }, + "unique_frac_abs_band": 0.9948052300347222, + "unique_frac_abs_band_std": 0.00021175801589619824, + "matched_frac_abs_per_threshold": { + "0.5": 0.7082248263888888, + "0.6": 0.6845160590277777, + "0.7": 0.6410590277777778, + "0.8": 0.5858289930555557, + "0.9": 0.5199110243055555 + }, + "matched_frac_abs_band": 0.6078287760416666, + "matched_frac_abs_band_std": 0.0035846573061908975, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/seed0/config.json b/vision/dinov2-b_layer10/relu/x8/l0_128/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..03af2764551fb0c668e41544170a64c51da927c1 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..5fc8d0553a2ee25d727e6a31a17718060a9c9322 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1317, + "r2_score": 0.9643, + "l0_per_token_0": 0.0208, + "l0_per_token_0.001": 0.0207, + "l0_per_token_0.01": 0.0202, + "l0_per_token_0.05": 0.0179, + "l0_per_token_0.1": 0.0154, + "l0_per_token_0.5": 0.0059, + "l0_per_token_1.0": 0.0028, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 1.0, + "l1_per_token": 74.5924 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/seed1/config.json b/vision/dinov2-b_layer10/relu/x8/l0_128/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2df92766c911a9cf15a36365f5f0d91bde86a2e6 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..20940c73b2f3309bce94a95dfb33bc51eec24fe4 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1319, + "r2_score": 0.9643, + "l0_per_token_0": 0.0207, + "l0_per_token_0.001": 0.0206, + "l0_per_token_0.01": 0.02, + "l0_per_token_0.05": 0.0177, + "l0_per_token_0.1": 0.0153, + "l0_per_token_0.5": 0.0059, + "l0_per_token_1.0": 0.0028, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 1.0, + "l1_per_token": 74.3676 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/seed2/config.json b/vision/dinov2-b_layer10/relu/x8/l0_128/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..6ff526fe86b1fee16310adf1b05c81222284af8c --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..abb2a0e2f30b788339505227e9c33534677e2b20 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_128/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.1312, + "r2_score": 0.9645, + "l0_per_token_0": 0.0208, + "l0_per_token_0.001": 0.0207, + "l0_per_token_0.01": 0.0202, + "l0_per_token_0.05": 0.0179, + "l0_per_token_0.1": 0.0154, + "l0_per_token_0.5": 0.0059, + "l0_per_token_1.0": 0.0028, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 1.0, + "l1_per_token": 74.616 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/code_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x8/l0_32/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..ef223d8e30a50aac56ec780e9241caac82485b00 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 32.58334241245136, + "recon_r2": 0.9377494054287051, + "l1_per_token": 44.55455462312277 + }, + "seed1": { + "l0": 32.022728923476, + "recon_r2": 0.9371990060330001, + "l1_per_token": 44.41416698506577 + }, + "seed2": { + "l0": 32.385273670557716, + "recon_r2": 0.9378435006268623, + "l1_per_token": 44.55453980001289 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.49376906290897943, + "mcc_abs": 0.4937862339858699, + "act_iou": 0.3381843737852636, + "act_iou_n_pairs": 6144, + "input_r2": 0.9942380390713804, + "matched_mcc_signed": { + "0.6": 0.909097006738297, + "0.7": 0.9292642756601239, + "0.8": 0.9466629776398878, + "0.9": 0.9702195534829933 + }, + "matched_mcc_abs": { + "0.6": 0.9082967320638157, + "0.7": 0.9286407024944627, + "0.8": 0.9465381480892502, + "0.9": 0.9702740386257488 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.4968002116420416, + "mcc_abs": 0.49681689687806374, + "act_iou": 0.3467555005312322, + "act_iou_n_pairs": 6144, + "input_r2": 0.9944229081839271, + "matched_mcc_signed": { + "0.6": 0.9097951855369142, + "0.7": 0.9285164209119564, + "0.8": 0.949635236477103, + "0.9": 0.9700750621080897 + }, + "matched_mcc_abs": { + "0.6": 0.9098846998320008, + "0.7": 0.928508402004818, + "0.8": 0.9496415448245251, + "0.9": 0.9700091491249002 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.4902599013888626, + "mcc_abs": 0.49027276412612797, + "act_iou": 0.3418821978670559, + "act_iou_n_pairs": 6144, + "input_r2": 0.9945830183781786, + "matched_mcc_signed": { + "0.6": 0.9118293633800736, + "0.7": 0.9323751444745814, + "0.8": 0.9510892064515788, + "0.9": 0.971793610334946 + }, + "matched_mcc_abs": { + "0.6": 0.9117882809693915, + "0.7": 0.9322559205436182, + "0.8": 0.9511196540052416, + "0.9": 0.9717262947538532 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 32.33044833549503, + "l0_std": 0.28429954982172, + "recon_r2_mean": 0.9375973040295226, + "recon_r2_std": 0.00034812992848126965, + "l1_per_token_mean": 44.50775380273381, + "l1_per_token_std": 0.08104856189874306, + "input_r2_mean": 0.994414655211162, + "input_r2_std": 0.00017263766738270255, + "mcc_signed_mean": 0.49360972531329456, + "mcc_signed_std": 0.0032730652153613554, + "mcc_abs_mean": 0.4936252983300205, + "mcc_abs_std": 0.0032750333712394998, + "matched_mcc_signed_per_threshold": { + "0.6": 0.9102405185517616, + "0.7": 0.930051947015554, + "0.8": 0.9491291401895232, + "0.9": 0.9706960753086763 + }, + "matched_mcc_signed_band": 0.9400294202663788, + "matched_mcc_abs_per_threshold": { + "0.6": 0.9099899042884027, + "0.7": 0.9298016750142996, + "0.8": 0.9490997823063388, + "0.9": 0.9706698275015008 + }, + "matched_mcc_abs_band": 0.9398902972776354, + "act_iou_mean": 0.3422740240611839, + "act_iou_std": 0.00429897653502182, + "act_iou_n_pairs": 6144.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/dict_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x8/l0_32/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..2c2b3bd65066bbf6a3afa14220f3c9fff80fc99d --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/dict_metrics_step_final.json @@ -0,0 +1,545 @@ +{ + "schema": 2, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 6144, + "d_in": 768, + "l0": 33, + "l0_raw": 32.5632, + "recon_r2": 0.9377, + "l1_per_token": 44.5719, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 6, + "dense_rate": 0.0009765625, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 687, + "0.6": 236, + "0.7": 72, + "0.8": 42, + "0.9": 25 + }, + "n_dup_clusters_signed": { + "0.5": 329, + "0.6": 112, + "0.7": 33, + "0.8": 20, + "0.9": 12 + }, + "largest_dup_signed": { + "0.5": 6, + "0.6": 6, + "0.7": 4, + "0.8": 4, + "0.9": 3 + }, + "babel_mu1_signed": 0.33007654547691345, + "n_dup_abs": { + "0.5": 692, + "0.6": 242, + "0.7": 73, + "0.8": 44, + "0.9": 25 + }, + "n_dup_clusters_abs": { + "0.5": 330, + "0.6": 114, + "0.7": 33, + "0.8": 21, + "0.9": 12 + }, + "largest_dup_abs": { + "0.5": 7, + "0.6": 7, + "0.7": 5, + "0.8": 4, + "0.9": 3 + }, + "babel_mu1_abs": 0.3317762613296509 + }, + "seed1": { + "d_sae": 6144, + "d_in": 768, + "l0": 32, + "l0_raw": 31.9488, + "recon_r2": 0.9372, + "l1_per_token": 44.4497, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 6, + "dense_rate": 0.0009765625, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 667, + "0.6": 227, + "0.7": 87, + "0.8": 53, + "0.9": 25 + }, + "n_dup_clusters_signed": { + "0.5": 317, + "0.6": 107, + "0.7": 38, + "0.8": 23, + "0.9": 10 + }, + "largest_dup_signed": { + "0.5": 10, + "0.6": 9, + "0.7": 8, + "0.8": 7, + "0.9": 7 + }, + "babel_mu1_signed": 0.32404109835624695, + "n_dup_abs": { + "0.5": 667, + "0.6": 230, + "0.7": 89, + "0.8": 55, + "0.9": 25 + }, + "n_dup_clusters_abs": { + "0.5": 316, + "0.6": 108, + "0.7": 39, + "0.8": 24, + "0.9": 10 + }, + "largest_dup_abs": { + "0.5": 11, + "0.6": 9, + "0.7": 8, + "0.8": 7, + "0.9": 7 + }, + "babel_mu1_abs": 0.32579511404037476 + }, + "seed2": { + "d_sae": 6144, + "d_in": 768, + "l0": 33, + "l0_raw": 32.5632, + "recon_r2": 0.9378, + "l1_per_token": 44.5604, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 6, + "dense_rate": 0.0009765625, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 641, + "0.6": 199, + "0.7": 62, + "0.8": 28, + "0.9": 16 + }, + "n_dup_clusters_signed": { + "0.5": 303, + "0.6": 91, + "0.7": 25, + "0.8": 11, + "0.9": 5 + }, + "largest_dup_signed": { + "0.5": 11, + "0.6": 9, + "0.7": 9, + "0.8": 8, + "0.9": 8 + }, + "babel_mu1_signed": 0.322643518447876, + "n_dup_abs": { + "0.5": 646, + "0.6": 198, + "0.7": 66, + "0.8": 32, + "0.9": 18 + }, + "n_dup_clusters_abs": { + "0.5": 305, + "0.6": 89, + "0.7": 27, + "0.8": 13, + "0.9": 6 + }, + "largest_dup_abs": { + "0.5": 10, + "0.6": 9, + "0.7": 9, + "0.8": 8, + "0.9": 8 + }, + "babel_mu1_abs": 0.324220210313797 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.5686493515968323, + "dcs_abs": 0.5709418654441833, + "frac_sign_flipped": 0.17805989583333334, + "alive_source": "stats", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.5576444268226624, + "0.6": 0.56437087059021, + "0.7": 0.5676257014274597, + "0.8": 0.5679740905761719, + "0.9": 0.5679279565811157 + }, + "unique_frac_signed": { + "0.5": 0.9423828125, + "0.6": 0.9801432291666666, + "0.7": 0.9928385416666666, + "0.8": 0.9957682291666666, + "0.9": 0.9977213541666666 + }, + "dedup_dcs_signed": { + "0.5": 0.525514523324091, + "0.6": 0.5531642875478913, + "0.7": 0.5635606736177579, + "0.8": 0.5655705543855826, + "0.9": 0.5666338499092186 + }, + "matched_frac_signed": { + "0.5": 0.4967447916666667, + "0.6": 0.50927734375, + "0.7": 0.4951171875, + "0.8": 0.4685872395833333, + "0.9": 0.40771484375 + }, + "matched_cos_abs": { + "0.5": 0.5599888563156128, + "0.6": 0.5665380954742432, + "0.7": 0.5698511600494385, + "0.8": 0.5702142715454102, + "0.9": 0.570227861404419 + }, + "unique_frac_abs": { + "0.5": 0.9419759114583334, + "0.6": 0.9796549479166666, + "0.7": 0.99267578125, + "0.8": 0.99560546875, + "0.9": 0.9977213541666666 + }, + "dedup_dcs_abs": { + "0.5": 0.5274960133344091, + "0.6": 0.5550118484146271, + "0.7": 0.5656774454982951, + "0.8": 0.5677084471099079, + "0.9": 0.5689285140639792 + }, + "matched_frac_abs": { + "0.5": 0.4962565104166667, + "0.6": 0.5089518229166666, + "0.7": 0.4957682291666667, + "0.8": 0.4684244791666667, + "0.9": 0.4075520833333333 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.5722612738609314, + "dcs_abs": 0.5746484994888306, + "frac_sign_flipped": 0.17350260416666666, + "alive_source": "stats", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.5625412464141846, + "0.6": 0.569828987121582, + "0.7": 0.5717218518257141, + "0.8": 0.5721675157546997, + "0.9": 0.5719816088676453 + }, + "unique_frac_signed": { + "0.5": 0.943359375, + "0.6": 0.9811197916666666, + "0.7": 0.9938151041666666, + "0.8": 0.996826171875, + "0.9": 0.998046875 + }, + "dedup_dcs_signed": { + "0.5": 0.5306785586290061, + "0.6": 0.5590704971303542, + "0.7": 0.5681858117265316, + "0.8": 0.5703515544009861, + "0.9": 0.5708644572878256 + }, + "matched_frac_signed": { + "0.5": 0.5026041666666666, + "0.6": 0.5157877604166666, + "0.7": 0.5013020833333334, + "0.8": 0.470703125, + "0.9": 0.421875 + }, + "matched_cos_abs": { + "0.5": 0.5650092363357544, + "0.6": 0.5723147392272949, + "0.7": 0.5739871859550476, + "0.8": 0.5744633078575134, + "0.9": 0.5743231177330017 + }, + "unique_frac_abs": { + "0.5": 0.9427897135416666, + "0.6": 0.980712890625, + "0.7": 0.9935709635416666, + "0.8": 0.99658203125, + "0.9": 0.9979654947916666 + }, + "dedup_dcs_abs": { + "0.5": 0.5326848960733817, + "0.6": 0.5612764422548935, + "0.7": 0.5702970014099265, + "0.8": 0.5724998102232348, + "0.9": 0.5731546543587076 + }, + "matched_frac_abs": { + "0.5": 0.5021158854166666, + "0.6": 0.5154622395833334, + "0.7": 0.5011393229166666, + "0.8": 0.470703125, + "0.9": 0.4220377604166667 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.566953718662262, + "dcs_abs": 0.569240152835846, + "frac_sign_flipped": 0.17008463541666666, + "alive_source": "stats", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.55628502368927, + "0.6": 0.5637752413749695, + "0.7": 0.566025972366333, + "0.8": 0.566567599773407, + "0.9": 0.5662784576416016 + }, + "unique_frac_signed": { + "0.5": 0.9440104166666666, + "0.6": 0.9814453125, + "0.7": 0.9930013020833334, + "0.8": 0.9961751302083334, + "0.9": 0.9978841145833334 + }, + "dedup_dcs_signed": { + "0.5": 0.5251388569983343, + "0.6": 0.5533145679510199, + "0.7": 0.5620645275727536, + "0.8": 0.5644005524760967, + "0.9": 0.5650802773113053 + }, + "matched_frac_signed": { + "0.5": 0.4928385416666667, + "0.6": 0.5066731770833334, + "0.7": 0.490234375, + "0.8": 0.4646809895833333, + "0.9": 0.4132486979166667 + }, + "matched_cos_abs": { + "0.5": 0.5585824847221375, + "0.6": 0.566035270690918, + "0.7": 0.5682031512260437, + "0.8": 0.5687363147735596, + "0.9": 0.568501889705658 + }, + "unique_frac_abs": { + "0.5": 0.9436848958333334, + "0.6": 0.981201171875, + "0.7": 0.9927571614583334, + "0.8": 0.9959309895833334, + "0.9": 0.997802734375 + }, + "dedup_dcs_abs": { + "0.5": 0.5271258539093349, + "0.6": 0.5553944709245116, + "0.7": 0.5640877475428473, + "0.8": 0.5664221207844093, + "0.9": 0.5672527400456602 + }, + "matched_frac_abs": { + "0.5": 0.4923502604166667, + "0.6": 0.5065104166666666, + "0.7": 0.48974609375, + "0.8": 0.4641927083333333, + "0.9": 0.4130859375 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 6, + "n_dense_std": 0.0, + "l0_mean": 33, + "l0_std": 0.35472400539010807, + "recon_r2_mean": 0.9375666666666667, + "recon_r2_std": 0.00032145502536645975, + "l1_per_token_mean": 44.52733333333333, + "l1_per_token_std": 0.06767498306858635, + "babel_mu1_signed_mean": 0.3255870540936788, + "babel_mu1_signed_std": 0.003950311008031424, + "n_dup_signed_per_threshold": { + "0.5": 665.0, + "0.6": 220.66666666666666, + "0.7": 73.66666666666667, + "0.8": 41.0, + "0.9": 22.0 + }, + "n_dup_signed_band": 89.33333333333333, + "n_dup_signed_band_std": 11.528045512285825, + "n_dup_clusters_signed_per_threshold": { + "0.5": 316.3333333333333, + "0.6": 103.33333333333333, + "0.7": 32.0, + "0.8": 18.0, + "0.9": 9.0 + }, + "n_dup_clusters_signed_band": 40.58333333333333, + "n_dup_clusters_signed_band_std": 6.568548799646185, + "largest_dup_signed_per_threshold": { + "0.5": 9.0, + "0.6": 8.0, + "0.7": 7.0, + "0.8": 6.333333333333333, + "0.9": 6.0 + }, + "largest_dup_signed_band": 6.833333333333333, + "largest_dup_signed_band_std": 2.268442931469367, + "dcs_signed_mean": 0.5692881147066752, + "dcs_signed_std": 0.0027108207760017807, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.5271106463171438, + "0.6": 0.5551831175430885, + "0.7": 0.5646036709723476, + "0.8": 0.5667742204208884, + "0.9": 0.5675261948361164 + }, + "dedup_dcs_signed_band": 0.5635218009431102, + "dedup_dcs_signed_band_std": 0.003155736584156142, + "matched_cos_signed_per_threshold": { + "0.5": 0.5588235656420389, + "0.6": 0.5659916996955872, + "0.7": 0.568457841873169, + "0.8": 0.5689030687014262, + "0.9": 0.5687293410301208 + }, + "matched_cos_signed_band": 0.5680204878250757, + "matched_cos_signed_band_std": 0.0030205738034055156, + "unique_frac_signed_per_threshold": { + "0.5": 0.9432508680555555, + "0.6": 0.9809027777777777, + "0.7": 0.9932183159722222, + "0.8": 0.9962565104166666, + "0.9": 0.9978841145833334 + }, + "unique_frac_signed_band": 0.9920654296875, + "unique_frac_signed_band_std": 0.0004204097151646223, + "matched_frac_signed_per_threshold": { + "0.5": 0.4973958333333333, + "0.6": 0.5105794270833334, + "0.7": 0.49555121527777785, + "0.8": 0.46799045138888884, + "0.9": 0.4142795138888889 + }, + "matched_frac_signed_band": 0.4721001519097222, + "matched_frac_signed_band_std": 0.00466240653606359, + "babel_mu1_abs_mean": 0.32726386189460754, + "babel_mu1_abs_std": 0.003986400874602514, + "n_dup_abs_per_threshold": { + "0.5": 668.3333333333334, + "0.6": 223.33333333333334, + "0.7": 76.0, + "0.8": 43.666666666666664, + "0.9": 22.666666666666668 + }, + "n_dup_abs_band": 91.41666666666669, + "n_dup_abs_band_std": 11.342214657346833, + "n_dup_clusters_abs_per_threshold": { + "0.5": 317.0, + "0.6": 103.66666666666667, + "0.7": 33.0, + "0.8": 19.333333333333332, + "0.9": 9.333333333333334 + }, + "n_dup_clusters_abs_band": 41.33333333333334, + "n_dup_clusters_abs_band_std": 6.568548799646185, + "largest_dup_abs_per_threshold": { + "0.5": 9.333333333333334, + "0.6": 8.333333333333334, + "0.7": 7.333333333333333, + "0.8": 6.333333333333333, + "0.9": 6.0 + }, + "largest_dup_abs_band": 7.0, + "largest_dup_abs_band_std": 1.984313483298443, + "dcs_abs_mean": 0.57161017258962, + "dcs_abs_std": 0.0027654166431468692, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.5291022544390419, + "0.6": 0.5572275871980107, + "0.7": 0.5666873981503563, + "0.8": 0.5688767927058507, + "0.9": 0.5697786361561157 + }, + "dedup_dcs_abs_band": 0.5656426035525833, + "dedup_dcs_abs_band_std": 0.003215947617554285, + "matched_cos_abs_per_threshold": { + "0.5": 0.5611935257911682, + "0.6": 0.5682960351308187, + "0.7": 0.5706804990768433, + "0.8": 0.5711379647254944, + "0.9": 0.5710176229476929 + }, + "matched_cos_abs_band": 0.5702830304702123, + "matched_cos_abs_band_std": 0.0030948607893854234, + "unique_frac_abs_per_threshold": { + "0.5": 0.9428168402777778, + "0.6": 0.9805230034722222, + "0.7": 0.9930013020833334, + "0.8": 0.9960394965277778, + "0.9": 0.997829861111111 + }, + "unique_frac_abs_band": 0.9918484157986112, + "unique_frac_abs_band_std": 0.0004019542511050876, + "matched_frac_abs_per_threshold": { + "0.5": 0.4969075520833333, + "0.6": 0.5103081597222222, + "0.7": 0.49555121527777773, + "0.8": 0.4677734375, + "0.9": 0.4142252604166667 + }, + "matched_frac_abs_band": 0.4719645182291667, + "matched_frac_abs_band_std": 0.00473685944363779, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/seed0/config.json b/vision/dinov2-b_layer10/relu/x8/l0_32/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..03af2764551fb0c668e41544170a64c51da927c1 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..8d1513bf837b826d84b199a0273460d7e522689a --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2299, + "r2_score": 0.9377, + "l0_per_token_0": 0.0053, + "l0_per_token_0.001": 0.0053, + "l0_per_token_0.01": 0.0052, + "l0_per_token_0.05": 0.005, + "l0_per_token_0.1": 0.0047, + "l0_per_token_0.5": 0.003, + "l0_per_token_1.0": 0.0019, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 0.999, + "l1_per_token": 44.5719 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/seed1/config.json b/vision/dinov2-b_layer10/relu/x8/l0_32/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2df92766c911a9cf15a36365f5f0d91bde86a2e6 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..c3a55ad2b7dfcaaea93c809621f07dc3d83cb7b0 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2321, + "r2_score": 0.9372, + "l0_per_token_0": 0.0052, + "l0_per_token_0.001": 0.0052, + "l0_per_token_0.01": 0.0052, + "l0_per_token_0.05": 0.0049, + "l0_per_token_0.1": 0.0046, + "l0_per_token_0.5": 0.003, + "l0_per_token_1.0": 0.0019, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 0.9998, + "l1_per_token": 44.4497 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/seed2/config.json b/vision/dinov2-b_layer10/relu/x8/l0_32/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..6ff526fe86b1fee16310adf1b05c81222284af8c --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..c4fa2eae415387e193700e97f3dd54eabedac9ad --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_32/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,15 @@ +{ + "l2_loss": 0.2297, + "r2_score": 0.9378, + "l0_per_token_0": 0.0053, + "l0_per_token_0.001": 0.0053, + "l0_per_token_0.01": 0.0052, + "l0_per_token_0.05": 0.0049, + "l0_per_token_0.1": 0.0046, + "l0_per_token_0.5": 0.003, + "l0_per_token_1.0": 0.0019, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0, + "live_latent_frac": 0.999, + "l1_per_token": 44.5604 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/code_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x8/l0_64/code_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..cb900312b411cad515be0acbd6ae63d114885555 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/code_metrics_step_final.json @@ -0,0 +1,136 @@ +{ + "schema": 1, + "step": "final", + "n_tokens": 771000, + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "l0": 54.887376134889756, + "recon_r2": 0.9490447819157598, + "l1_per_token": 52.349633696271226 + }, + "seed1": { + "l0": 53.645396887159535, + "recon_r2": 0.9490777437498327, + "l1_per_token": 52.25126665265867 + }, + "seed2": { + "l0": 54.23844617380026, + "recon_r2": 0.9491861120521087, + "l1_per_token": 52.27526610994597 + } + }, + "per_pair": { + "seed0\u2194seed1": { + "alive_source": "stats", + "mcc_signed": 0.661796807831626, + "mcc_abs": 0.6618132365607666, + "act_iou": 0.44432449550886893, + "act_iou_n_pairs": 6144, + "input_r2": 0.9944165290208118, + "matched_mcc_signed": { + "0.6": 0.8758737084491768, + "0.7": 0.9082881257698917, + "0.8": 0.9401731412257639, + "0.9": 0.967734421218839 + }, + "matched_mcc_abs": { + "0.6": 0.8755836471888763, + "0.7": 0.9083180062697745, + "0.8": 0.9398628193492148, + "0.9": 0.9679055938838963 + } + }, + "seed0\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6585044907544876, + "mcc_abs": 0.6585196270679695, + "act_iou": 0.4435362675078341, + "act_iou_n_pairs": 6144, + "input_r2": 0.9943807532731497, + "matched_mcc_signed": { + "0.6": 0.8730213921516602, + "0.7": 0.9070300014442769, + "0.8": 0.9391875605269219, + "0.9": 0.968312066657271 + }, + "matched_mcc_abs": { + "0.6": 0.8734023753008943, + "0.7": 0.9067467293363529, + "0.8": 0.9394364967851455, + "0.9": 0.9682945793900115 + } + }, + "seed1\u2194seed2": { + "alive_source": "stats", + "mcc_signed": 0.6563320946507327, + "mcc_abs": 0.6563441571974827, + "act_iou": 0.4462535942225068, + "act_iou_n_pairs": 6144, + "input_r2": 0.9947131994105504, + "matched_mcc_signed": { + "0.6": 0.8774830639262399, + "0.7": 0.9090806723335872, + "0.8": 0.9387216471619884, + "0.9": 0.9667599403577826 + }, + "matched_mcc_abs": { + "0.6": 0.8772138327264661, + "0.7": 0.9092146341981573, + "0.8": 0.9387581844902289, + "0.9": 0.9667677181079208 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "l0_mean": 54.25707306528319, + "l0_std": 0.6211991095960044, + "recon_r2_mean": 0.9491028792392338, + "recon_r2_std": 7.39418452614641e-05, + "l1_per_token_mean": 52.29205548629196, + "l1_per_token_std": 0.051287729388558266, + "input_r2_mean": 0.9945034939015039, + "input_r2_std": 0.00018248911316434636, + "mcc_signed_mean": 0.6588777977456154, + "mcc_signed_std": 0.002751416202535671, + "mcc_abs_mean": 0.6588923402754063, + "mcc_abs_std": 0.002753523855303017, + "matched_mcc_signed_per_threshold": { + "0.6": 0.8754593881756924, + "0.7": 0.9081329331825853, + "0.8": 0.9393607829715581, + "0.9": 0.9676021427446307 + }, + "matched_mcc_signed_band": 0.9226388117686166, + "matched_mcc_abs_per_threshold": { + "0.6": 0.8753999517387455, + "0.7": 0.908093123268095, + "0.8": 0.9393525002081965, + "0.9": 0.9676559637939429 + }, + "matched_mcc_abs_band": 0.922625384752245, + "act_iou_mean": 0.44470478574640326, + "act_iou_std": 0.0013980098773404542, + "act_iou_n_pairs": 6144.0, + "alive_source": "stats" + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/dict_metrics_step_final.json b/vision/dinov2-b_layer10/relu/x8/l0_64/dict_metrics_step_final.json new file mode 100644 index 0000000000000000000000000000000000000000..ab8592a6f84ae440748b163f4cb07fe239d91d4f --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/dict_metrics_step_final.json @@ -0,0 +1,595 @@ +{ + "schema": 3, + "step": "final", + "linkage": "complete", + "thresholds": [ + 0.5, + 0.6, + 0.7, + 0.8, + 0.9 + ], + "seeds": [ + "seed0", + "seed1", + "seed2" + ], + "per_seed": { + "seed0": { + "d_sae": 6144, + "d_in": 768, + "l0": 55, + "l0_raw": 54.6816, + "recon_r2": 0.9513, + "l1_per_token": 52.3621, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 5, + "dense_rate": 0.0008138020833333334, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 1082, + "0.6": 351, + "0.7": 132, + "0.8": 67, + "0.9": 21 + }, + "n_dup_clusters_signed": { + "0.5": 522, + "0.6": 169, + "0.7": 62, + "0.8": 32, + "0.9": 9 + }, + "n_dup_removed_signed": { + "0.5": 560, + "0.6": 182, + "0.7": 70, + "0.8": 35, + "0.9": 12 + }, + "largest_dup_signed": { + "0.5": 8, + "0.6": 7, + "0.7": 7, + "0.8": 5, + "0.9": 5 + }, + "babel_mu1_signed": 0.40699079632759094, + "n_dup_abs": { + "0.5": 1095, + "0.6": 356, + "0.7": 132, + "0.8": 69, + "0.9": 23 + }, + "n_dup_clusters_abs": { + "0.5": 527, + "0.6": 171, + "0.7": 62, + "0.8": 33, + "0.9": 10 + }, + "n_dup_removed_abs": { + "0.5": 568, + "0.6": 185, + "0.7": 70, + "0.8": 36, + "0.9": 13 + }, + "largest_dup_abs": { + "0.5": 11, + "0.6": 7, + "0.7": 7, + "0.8": 5, + "0.9": 5 + }, + "babel_mu1_abs": 0.40848979353904724, + "dup_excludes_dead": true + }, + "seed1": { + "d_sae": 6144, + "d_in": 768, + "l0": 53, + "l0_raw": 53.452799999999996, + "recon_r2": 0.9513, + "l1_per_token": 52.266, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 2, + "dense_rate": 0.0003255208333333333, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 1055, + "0.6": 371, + "0.7": 136, + "0.8": 71, + "0.9": 33 + }, + "n_dup_clusters_signed": { + "0.5": 504, + "0.6": 176, + "0.7": 63, + "0.8": 32, + "0.9": 14 + }, + "n_dup_removed_signed": { + "0.5": 551, + "0.6": 195, + "0.7": 73, + "0.8": 39, + "0.9": 19 + }, + "largest_dup_signed": { + "0.5": 9, + "0.6": 9, + "0.7": 8, + "0.8": 7, + "0.9": 6 + }, + "babel_mu1_signed": 0.40160369873046875, + "n_dup_abs": { + "0.5": 1063, + "0.6": 376, + "0.7": 138, + "0.8": 73, + "0.9": 33 + }, + "n_dup_clusters_abs": { + "0.5": 506, + "0.6": 178, + "0.7": 64, + "0.8": 33, + "0.9": 14 + }, + "n_dup_removed_abs": { + "0.5": 557, + "0.6": 198, + "0.7": 74, + "0.8": 40, + "0.9": 19 + }, + "largest_dup_abs": { + "0.5": 11, + "0.6": 9, + "0.7": 8, + "0.8": 7, + "0.9": 6 + }, + "babel_mu1_abs": 0.40321967005729675, + "dup_excludes_dead": true + }, + "seed2": { + "d_sae": 6144, + "d_in": 768, + "l0": 54, + "l0_raw": 54.0672, + "recon_r2": 0.9514, + "l1_per_token": 52.2746, + "stats_source": "stats", + "n_dead": 0, + "dead_rate": 0.0, + "alive_source": "stats", + "n_dense": 4, + "dense_rate": 0.0006510416666666666, + "dense_threshold": 0.1, + "n_dup_signed": { + "0.5": 1021, + "0.6": 360, + "0.7": 137, + "0.8": 63, + "0.9": 30 + }, + "n_dup_clusters_signed": { + "0.5": 489, + "0.6": 173, + "0.7": 64, + "0.8": 29, + "0.9": 13 + }, + "n_dup_removed_signed": { + "0.5": 532, + "0.6": 187, + "0.7": 73, + "0.8": 34, + "0.9": 17 + }, + "largest_dup_signed": { + "0.5": 10, + "0.6": 8, + "0.7": 8, + "0.8": 7, + "0.9": 6 + }, + "babel_mu1_signed": 0.40266790986061096, + "n_dup_abs": { + "0.5": 1027, + "0.6": 367, + "0.7": 136, + "0.8": 65, + "0.9": 30 + }, + "n_dup_clusters_abs": { + "0.5": 492, + "0.6": 175, + "0.7": 63, + "0.8": 30, + "0.9": 13 + }, + "n_dup_removed_abs": { + "0.5": 535, + "0.6": 192, + "0.7": 73, + "0.8": 35, + "0.9": 17 + }, + "largest_dup_abs": { + "0.5": 11, + "0.6": 8, + "0.7": 8, + "0.8": 7, + "0.9": 6 + }, + "babel_mu1_abs": 0.404160737991333, + "dup_excludes_dead": true + } + }, + "per_pair": { + "seed0\u2194seed1": { + "dcs_signed": 0.7374071478843689, + "dcs_abs": 0.737889289855957, + "frac_sign_flipped": 0.033040364583333336, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.7349286079406738, + "0.6": 0.7379416823387146, + "0.7": 0.7375900745391846, + "0.8": 0.7371759414672852, + "0.9": 0.7372441291809082 + }, + "unique_frac_signed": { + "0.5": 0.9095865885416666, + "0.6": 0.9693196614583334, + "0.7": 0.9883626302083334, + "0.8": 0.9939778645833334, + "0.9": 0.9974772135416666 + }, + "dedup_dcs_signed": { + "0.5": 0.6684812053184335, + "0.6": 0.7153013817005558, + "0.7": 0.7290064660871092, + "0.8": 0.7327365681218605, + "0.9": 0.7353842196753249 + }, + "matched_frac_signed": { + "0.5": 0.685546875, + "0.6": 0.6861979166666666, + "0.7": 0.6481119791666666, + "0.8": 0.5826822916666666, + "0.9": 0.49755859375 + }, + "matched_cos_abs": { + "0.5": 0.7354100346565247, + "0.6": 0.7383738160133362, + "0.7": 0.7381297945976257, + "0.8": 0.7375937104225159, + "0.9": 0.7376465797424316 + }, + "unique_frac_abs": { + "0.5": 0.908447265625, + "0.6": 0.9688313802083334, + "0.7": 0.98828125, + "0.8": 0.9938151041666666, + "0.9": 0.9973958333333334 + }, + "dedup_dcs_abs": { + "0.5": 0.6680812350969063, + "0.6": 0.7153597232778945, + "0.7": 0.7294798360671848, + "0.8": 0.7330317701562308, + "0.9": 0.7357256251076858 + }, + "matched_frac_abs": { + "0.5": 0.6845703125, + "0.6": 0.6863606770833334, + "0.7": 0.6477864583333334, + "0.8": 0.5830078125, + "0.9": 0.4973958333333333 + } + }, + "seed0\u2194seed2": { + "dcs_signed": 0.7360313534736633, + "dcs_abs": 0.7365620732307434, + "frac_sign_flipped": 0.033365885416666664, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.7364463210105896, + "0.6": 0.7371519207954407, + "0.7": 0.7368056178092957, + "0.8": 0.736416757106781, + "0.9": 0.7362435460090637 + }, + "unique_frac_signed": { + "0.5": 0.9111328125, + "0.6": 0.969970703125, + "0.7": 0.9883626302083334, + "0.8": 0.994384765625, + "0.9": 0.9976399739583334 + }, + "dedup_dcs_signed": { + "0.5": 0.6710004077176563, + "0.6": 0.7150157669238979, + "0.7": 0.7282311383702715, + "0.8": 0.732281604417949, + "0.9": 0.7345059920674734 + }, + "matched_frac_signed": { + "0.5": 0.681640625, + "0.6": 0.6853841145833334, + "0.7": 0.6442057291666666, + "0.8": 0.5812174479166666, + "0.9": 0.4925130208333333 + }, + "matched_cos_abs": { + "0.5": 0.7372598648071289, + "0.6": 0.7377150058746338, + "0.7": 0.7373217344284058, + "0.8": 0.7369011640548706, + "0.9": 0.7366864085197449 + }, + "unique_frac_abs": { + "0.5": 0.9102376302083334, + "0.6": 0.9693196614583334, + "0.7": 0.9883626302083334, + "0.8": 0.9942220052083334, + "0.9": 0.99755859375 + }, + "dedup_dcs_abs": { + "0.5": 0.6710816721897572, + "0.6": 0.7150816597471324, + "0.7": 0.7287412487494294, + "0.8": 0.7326433529669885, + "0.9": 0.7348878577176947 + }, + "matched_frac_abs": { + "0.5": 0.6805013020833334, + "0.6": 0.6847330729166666, + "0.7": 0.64453125, + "0.8": 0.5807291666666666, + "0.9": 0.4923502604166667 + } + }, + "seed1\u2194seed2": { + "dcs_signed": 0.7342228889465332, + "dcs_abs": 0.7348442673683167, + "frac_sign_flipped": 0.0390625, + "alive_source": "stats", + "frac_denominator": "d_sae", + "n_alive_a": 6144, + "n_alive_b": 6144, + "matched_cos_signed": { + "0.5": 0.7331328392028809, + "0.6": 0.7351481914520264, + "0.7": 0.7348409295082092, + "0.8": 0.7344930768013, + "0.9": 0.7338583469390869 + }, + "unique_frac_signed": { + "0.5": 0.911865234375, + "0.6": 0.9689127604166666, + "0.7": 0.9881184895833334, + "0.8": 0.9940592447916666, + "0.9": 0.9970703125 + }, + "dedup_dcs_signed": { + "0.5": 0.6685183482477441, + "0.6": 0.712294463495103, + "0.7": 0.7261099093496645, + "0.8": 0.730129633229808, + "0.9": 0.7317083713132888 + }, + "matched_frac_signed": { + "0.5": 0.6795247395833334, + "0.6": 0.6832682291666666, + "0.7": 0.64404296875, + "0.8": 0.5849609375, + "0.9": 0.50048828125 + }, + "matched_cos_abs": { + "0.5": 0.7338494658470154, + "0.6": 0.7355476021766663, + "0.7": 0.7355025410652161, + "0.8": 0.7350829243659973, + "0.9": 0.7344829440116882 + }, + "unique_frac_abs": { + "0.5": 0.9111328125, + "0.6": 0.96826171875, + "0.7": 0.988037109375, + "0.8": 0.993896484375, + "0.9": 0.9970703125 + }, + "dedup_dcs_abs": { + "0.5": 0.6686343277688138, + "0.6": 0.7122025855060201, + "0.7": 0.7267038046120433, + "0.8": 0.7305963342514588, + "0.9": 0.732331138511654 + }, + "matched_frac_abs": { + "0.5": 0.6785481770833334, + "0.6": 0.6829427083333334, + "0.7": 0.6435546875, + "0.8": 0.5849609375, + "0.9": 0.50048828125 + } + } + }, + "summary": { + "n_seeds": 3, + "n_pairs": 3, + "report_band": [ + 0.6, + 0.7, + 0.8, + 0.9 + ], + "n_dead_mean": 0, + "n_dead_std": 0.0, + "n_dense_mean": 4, + "n_dense_std": 1.5275252316519468, + "l0_mean": 54, + "l0_std": 0.6144000000000034, + "recon_r2_mean": 0.9513333333333334, + "recon_r2_std": 5.773502691895622e-05, + "l1_per_token_mean": 52.30089999999999, + "l1_per_token_std": 0.0531749000939349, + "babel_mu1_signed_mean": 0.4037541349728902, + "babel_mu1_signed_std": 0.0028530893564940245, + "n_dup_signed_per_threshold": { + "0.5": 1052.6666666666667, + "0.6": 360.6666666666667, + "0.7": 135.0, + "0.8": 67.0, + "0.9": 28.0 + }, + "n_dup_signed_band": 147.66666666666669, + "n_dup_signed_band_std": 5.0020828994863065, + "n_dup_clusters_signed_per_threshold": { + "0.5": 505.0, + "0.6": 172.66666666666666, + "0.7": 63.0, + "0.8": 31.0, + "0.9": 12.0 + }, + "n_dup_clusters_signed_band": 69.66666666666666, + "n_dup_clusters_signed_band_std": 1.626601774661928, + "largest_dup_signed_per_threshold": { + "0.5": 9.0, + "0.6": 8.0, + "0.7": 7.666666666666667, + "0.8": 6.333333333333333, + "0.9": 5.666666666666667 + }, + "largest_dup_signed_band": 6.916666666666667, + "largest_dup_signed_band_std": 0.8036375634160796, + "dcs_signed_mean": 0.7358871301015218, + "dcs_signed_std": 0.001597021143128834, + "dedup_dcs_signed_per_threshold": { + "0.5": 0.6693333204279446, + "0.6": 0.7142038707065188, + "0.7": 0.7277825046023484, + "0.8": 0.7317159352565392, + "0.9": 0.733866194352029 + }, + "dedup_dcs_signed_band": 0.7268921262293588, + "dedup_dcs_signed_band_std": 0.001614138268628737, + "matched_cos_signed_per_threshold": { + "0.5": 0.7348359227180481, + "0.6": 0.7367472648620605, + "0.7": 0.7364122072855631, + "0.8": 0.7360285917917887, + "0.9": 0.7357820073763529 + }, + "matched_cos_signed_band": 0.7362425178289412, + "matched_cos_signed_band_std": 0.0014946118370145324, + "unique_frac_signed_per_threshold": { + "0.5": 0.9108615451388888, + "0.6": 0.9694010416666666, + "0.7": 0.98828125, + "0.8": 0.994140625, + "0.9": 0.9973958333333334 + }, + "unique_frac_signed_band": 0.9873046875, + "unique_frac_signed_band_std": 0.0002752227632338767, + "matched_frac_signed_per_threshold": { + "0.5": 0.6822374131944445, + "0.6": 0.6849500868055555, + "0.7": 0.6454535590277778, + "0.8": 0.5829535590277778, + "0.9": 0.4968532986111111 + }, + "matched_frac_signed_band": 0.6025526258680556, + "matched_frac_signed_band_std": 0.0015084636236205422, + "babel_mu1_abs_mean": 0.40529006719589233, + "babel_mu1_abs_std": 0.0028107096465888153, + "n_dup_abs_per_threshold": { + "0.5": 1061.6666666666667, + "0.6": 366.3333333333333, + "0.7": 135.33333333333334, + "0.8": 69.0, + "0.9": 28.666666666666668 + }, + "n_dup_abs_band": 149.83333333333331, + "n_dup_abs_band_std": 5.008326400438906, + "n_dup_clusters_abs_per_threshold": { + "0.5": 508.3333333333333, + "0.6": 174.66666666666666, + "0.7": 63.0, + "0.8": 32.0, + "0.9": 12.333333333333334 + }, + "n_dup_clusters_abs_band": 70.49999999999999, + "n_dup_clusters_abs_band_std": 1.6393596310755, + "largest_dup_abs_per_threshold": { + "0.5": 11.0, + "0.6": 8.0, + "0.7": 7.666666666666667, + "0.8": 6.333333333333333, + "0.9": 5.666666666666667 + }, + "largest_dup_abs_band": 6.916666666666667, + "largest_dup_abs_band_std": 0.8036375634160796, + "dcs_abs_mean": 0.736431876818339, + "dcs_abs_std": 0.001526680653218325, + "dedup_dcs_abs_per_threshold": { + "0.5": 0.6692657450184925, + "0.6": 0.7142146561770156, + "0.7": 0.7283082964762192, + "0.8": 0.7320904857915593, + "0.9": 0.7343148737790116 + }, + "dedup_dcs_abs_band": 0.7272320780559515, + "dedup_dcs_abs_band_std": 0.0015613693168047308, + "matched_cos_abs_per_threshold": { + "0.5": 0.7355064551035563, + "0.6": 0.7372121413548788, + "0.7": 0.7369846900304159, + "0.8": 0.7365259329477946, + "0.9": 0.7362719774246216 + }, + "matched_cos_abs_band": 0.7367486854394277, + "matched_cos_abs_band_std": 0.0014350327913320755, + "unique_frac_abs_per_threshold": { + "0.5": 0.9099392361111112, + "0.6": 0.9688042534722223, + "0.7": 0.9882269965277778, + "0.8": 0.9939778645833334, + "0.9": 0.9973415798611112 + }, + "unique_frac_abs_band": 0.9870876736111112, + "unique_frac_abs_band_std": 0.0002747209893192418, + "matched_frac_abs_per_threshold": { + "0.5": 0.6812065972222223, + "0.6": 0.6846788194444445, + "0.7": 0.6452907986111112, + "0.8": 0.5828993055555555, + "0.9": 0.4967447916666667 + }, + "matched_frac_abs_band": 0.6024034288194444, + "matched_frac_abs_band_std": 0.0016073020360611073, + "alive_source": "stats", + "frac_denominator": "d_sae", + "d_sae": 6144 + } +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/seed0/config.json b/vision/dinov2-b_layer10/relu/x8/l0_64/seed0/config.json new file mode 100644 index 0000000000000000000000000000000000000000..03af2764551fb0c668e41544170a64c51da927c1 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/seed0/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 0, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..d08fbeac5b840ae8724645c9c3b70fd0daed4407 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/seed0/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.1883, + "r2_score": 0.9513, + "l0_per_token_0": 0.0089, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 1.0, + "l1_per_token": 52.3621 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/seed1/config.json b/vision/dinov2-b_layer10/relu/x8/l0_64/seed1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2df92766c911a9cf15a36365f5f0d91bde86a2e6 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/seed1/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 1, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..6a7d48587e76884ce47fd7323f276b76dcca335f --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/seed1/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.1882, + "r2_score": 0.9513, + "l0_per_token_0": 0.0087, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 1.0, + "l1_per_token": 52.266 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/seed2/config.json b/vision/dinov2-b_layer10/relu/x8/l0_64/seed2/config.json new file mode 100644 index 0000000000000000000000000000000000000000..6ff526fe86b1fee16310adf1b05c81222284af8c --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/seed2/config.json @@ -0,0 +1,10 @@ +{ + "d_in": 768, + "d_sae_factor": 8, + "architecture": "relu", + "activation_name": "relu", + "device": "cpu", + "seed": 2, + "class_type": "VanillaSAE", + "module_name": "fastsae.models.sae" +} diff --git a/vision/dinov2-b_layer10/relu/x8/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json b/vision/dinov2-b_layer10/relu/x8/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json new file mode 100644 index 0000000000000000000000000000000000000000..024367bcf4a49c27598959a9ec0d3b8270ba4133 --- /dev/null +++ b/vision/dinov2-b_layer10/relu/x8/l0_64/seed2/stats.imagenet.split_val.seed_no/metrics.json @@ -0,0 +1,9 @@ +{ + "l2_loss": 0.1878, + "r2_score": 0.9514, + "l0_per_token_0": 0.0088, + "dead_rate": 0.0, + "dense_rate_0.2": 0.0002, + "live_latent_frac": 1.0, + "l1_per_token": 52.2746 +} \ No newline at end of file diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..e74e84ba1be89187883bcde4b2e7b1f66e9743e8 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e466da84ec4028ae762a49d2ef16d88e84b2406064615204274f9eb9edfbba5 +size 1600 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..7943e0c68ef98bf40b19537b3b4b187f3e3fe467 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec2d8e53f8001d0c4ee8d874612f6a4996e32dfe123dd35633309fa837cb93d3 +size 1502 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/firing_rate_per_latent.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/firing_rate_per_latent.pt new file mode 100644 index 0000000000000000000000000000000000000000..6c730501180907565c9e7ff0e4284ca62e8f08c0 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/firing_rate_per_latent.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c836a575c30069b359c7684db5952742c0c4df9845bff4ed26530f35b81c593 +size 50471 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/max_act_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/max_act_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..8c86779f2929eff62841e48ad4fbd859d04b1525 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/max_act_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba22f9ddf2e6793fd35e043cd1e5b5fa99809cec722b6b57e1b313cda2ec2a84 +size 50382 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/mean_act_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/mean_act_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..d0c29632aad857a1780ecd56d81fea2d732d3cc0 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/mean_act_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0b1074c28c366963570a8eb1f842c376e875516f9bbb58dfc540498f119a657 +size 50387 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/sparsity.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/sparsity.pt new file mode 100644 index 0000000000000000000000000000000000000000..ffd2acaa231df75224555b052d4ff380f77ebdf7 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/sparsity.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f20b6c6235748d5ca861a629ccb7c5482b2fd8a40ef58be5cff47f1ec4941c28 +size 50337 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/sparsity_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/sparsity_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..14a3e97eeb4d3508459caedc610edd93fd941239 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed1/stats.imagenet.split_val.seed_no/sparsity_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:230f043483de78a8e96ceb552891a439946e4338e916e99d24e98243750a3e35 +size 50387 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..e74e84ba1be89187883bcde4b2e7b1f66e9743e8 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e466da84ec4028ae762a49d2ef16d88e84b2406064615204274f9eb9edfbba5 +size 1600 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/sparsity.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/sparsity.pt new file mode 100644 index 0000000000000000000000000000000000000000..73ef50464f99595ddac602d6d4a1ac87126d3bfb --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/sparsity.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fc6bd856e352675dff242e957b7844b7c26e45efef7129627c8d7593159d59c +size 50337 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/sparsity_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/sparsity_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..53cc5295c0c9f5e804f97fc49ff301ebdf39f9d1 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_32/seed2/stats.imagenet.split_val.seed_no/sparsity_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:197c8743ca027e5cf07a4c2af3a788ee013c782403a88388b01201e839233815 +size 50387 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/model.safetensors b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/model.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..efb143abc794d9058ed6e6bcc859a18732db7f9d --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/model.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffa0dfb6aa46d56d71cd1c012d30028cbc14bb1f07e520b2e4d01548eb6a905f +size 75550008 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..35f74e97bfe3ef9a0cc0d6ecc07e7a7ab93a16c0 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/active_count_hist_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16b21dc49af69324aa3312244af41c90d7d044a3afbe24af9de70216391b5d55 +size 1856 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..547ba1231af9d8775a0caf42b3fbc232f8f9aee1 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99bad9e348112bb7d9d0a96cff4c8e67237e2389e0598597f778206fc09d42bb +size 1630 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/firing_rate_per_latent.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/firing_rate_per_latent.pt new file mode 100644 index 0000000000000000000000000000000000000000..dc249d47a4ab7f5051488dddc55c5f4b9b503c18 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/firing_rate_per_latent.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b6e007c979f1b913cfbb1348c8a5bdad69b6129454cef2e3ac0d462f473087d +size 50471 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/max_act_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/max_act_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..12b3998ab6afa0f49fe3753160533578d26ec5f2 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/max_act_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7efb7bd1b1a38fa46eee8b16aaa46679267d024a03b16fc77e756c31ec398898 +size 50382 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/mean_act.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/mean_act.pt new file mode 100644 index 0000000000000000000000000000000000000000..e6ab6bcd23d8b755ad6346eea3adc9b81296012d --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/mean_act.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c52fb2dd02ed25b1a6571729b3439c26e2dcba48ee5aceb003a013f63fc0848 +size 50337 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/mean_act_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/mean_act_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..c0aa94583ea28eb70bd0c7b038385e7a2588aadc --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/mean_act_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db0b021f97235b68921409907999a723cd2236382602c7c457597a81f4eed78b +size 50387 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/sparsity.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/sparsity.pt new file mode 100644 index 0000000000000000000000000000000000000000..fb2487e6d03091ae9d37f09bdb7e8aba8bdf662c --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/sparsity.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f3c92bd3e4a58f6f8e4b17cdc15e7d538b0948c44206978c4c169ae93a12403 +size 50337 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/sparsity_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/sparsity_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..d0d65ee838e992217cc136166e52d8051512c906 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed0/stats.imagenet.split_val.seed_no/sparsity_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a29fda52ab98bdc0fb110c73160e2f6baa0f6a8bcb661a4688f5dc2b581caba +size 50387 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed1/model.safetensors b/vision/dinov2-b_layer10/topk/x16/l0_64/seed1/model.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e2b146750c1ecba4baceea320e02d04a468df09c --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed1/model.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6e66be67c38b4f745cf29a52eb4ee6b1771d53b51bdd32b063dc3557dc7c89d +size 75550008 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/model.safetensors b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/model.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..829be9ba05a805393276667e36cfc0c0e1745f7e --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/model.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bec2bab6e9db8e8e314da6634d392a392017875688847fef834552a7f77972ae +size 75550008 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..547ba1231af9d8775a0caf42b3fbc232f8f9aee1 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/active_count_proportion_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99bad9e348112bb7d9d0a96cff4c8e67237e2389e0598597f778206fc09d42bb +size 1630 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/max_act_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/max_act_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..ed871c6e3e6caa27c1408e5f9c16064d910255b8 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/max_act_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53efdb1cc0e1dcd956a35b25640cb8aa486b443397288dd29db4d30e8f9608bb +size 50382 diff --git a/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/sparsity_per_token.pt b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/sparsity_per_token.pt new file mode 100644 index 0000000000000000000000000000000000000000..bf607d924ccb22b38c9877f396c0c59d8f8bd3d6 --- /dev/null +++ b/vision/dinov2-b_layer10/topk/x16/l0_64/seed2/stats.imagenet.split_val.seed_no/sparsity_per_token.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee6b338a8620b8da7b6b7a85deabc723e5d7840e0e60d5c7be98544e4ce576a0 +size 50387