Add files using upload-large-folder tool

lm/gpt2-small_layer8/batchtopk/x16/l0_128/code_metrics_step_final.json

@@ -0,0 +1,136 @@
+{
+ "schema": 1,
+ "step": "final",
+ "n_tokens": 600000,
+ "thresholds": [
+  0.5,
+  0.6,
+  0.7,
+  0.8,
+  0.9
+ ],
+ "seeds": [
+  "seed0",
+  "seed1",
+  "seed2"
+ ],
+ "per_seed": {
+  "seed0": {
+   "l0": 127.60186833333333,
+   "recon_r2": 0.9549164696200511,
+   "l1_per_token": 607.733303906817
+  },
+  "seed1": {
+   "l0": 128.26901333333333,
+   "recon_r2": 0.9550364366052586,
+   "l1_per_token": 610.3533042558479
+  },
+  "seed2": {
+   "l0": 127.33541666666666,
+   "recon_r2": 0.9548217346630481,
+   "l1_per_token": 606.1776392972702
+  }
+ },
+ "per_pair": {
+  "seed0\u2194seed1": {
+   "alive_source": "stats",
+   "mcc_signed": 0.7197823649387445,
+   "mcc_abs": 0.7198161701249538,
+   "act_iou": 0.3765977592859709,
+   "act_iou_n_pairs": 12288,
+   "input_r2": 0.9726636356267777,
+   "matched_mcc_signed": {
+    "0.6": 0.8604219640832733,
+    "0.7": 0.8902659821617671,
+    "0.8": 0.9180785324351199,
+    "0.9": 0.9493570893912988
+   },
+   "matched_mcc_abs": {
+    "0.6": 0.8604668676010798,
+    "0.7": 0.8905839594546318,
+    "0.8": 0.9183884360757144,
+    "0.9": 0.9498763732181995
+   }
+  },
+  "seed0\u2194seed2": {
+   "alive_source": "stats",
+   "mcc_signed": 0.7164556125489968,
+   "mcc_abs": 0.7164774781849763,
+   "act_iou": 0.37359442046385594,
+   "act_iou_n_pairs": 12288,
+   "input_r2": 0.9724552338792009,
+   "matched_mcc_signed": {
+    "0.6": 0.8593533231762739,
+    "0.7": 0.8896758148365826,
+    "0.8": 0.9198472560480646,
+    "0.9": 0.9504334646407947
+   },
+   "matched_mcc_abs": {
+    "0.6": 0.8599169115633304,
+    "0.7": 0.890333888426092,
+    "0.8": 0.9201968236292594,
+    "0.9": 0.9505123240419678
+   }
+  },
+  "seed1\u2194seed2": {
+   "alive_source": "stats",
+   "mcc_signed": 0.7184330381742603,
+   "mcc_abs": 0.7184667161459387,
+   "act_iou": 0.3755397594742718,
+   "act_iou_n_pairs": 12288,
+   "input_r2": 0.9727663797231461,
+   "matched_mcc_signed": {
+    "0.6": 0.8577225733053204,
+    "0.7": 0.8891815496720875,
+    "0.8": 0.9174585580331756,
+    "0.9": 0.9486653999848231
+   },
+   "matched_mcc_abs": {
+    "0.6": 0.8577840632335128,
+    "0.7": 0.8894980949335266,
+    "0.8": 0.9177318409381179,
+    "0.9": 0.9487187715656917
+   }
+  }
+ },
+ "summary": {
+  "n_seeds": 3,
+  "n_pairs": 3,
+  "report_band": [
+   0.6,
+   0.7,
+   0.8,
+   0.9
+  ],
+  "l0_mean": 127.73543277777777,
+  "l0_std": 0.48091608375848277,
+  "recon_r2_mean": 0.9549248802961192,
+  "recon_r2_std": 0.00010759779552195764,
+  "l1_per_token_mean": 608.088082486645,
+  "l1_per_token_std": 2.1103187773656535,
+  "input_r2_mean": 0.9726284164097082,
+  "input_r2_std": 0.00015853463025262014,
+  "mcc_signed_mean": 0.7182236718873338,
+  "mcc_signed_std": 0.00167322922734722,
+  "mcc_abs_mean": 0.718253454818623,
+  "mcc_abs_std": 0.0016795315605293117,
+  "matched_mcc_signed_per_threshold": {
+   "0.6": 0.8591659535216225,
+   "0.7": 0.889707782223479,
+   "0.8": 0.9184614488387867,
+   "0.9": 0.9494853180056388
+  },
+  "matched_mcc_signed_band": 0.9042051256473818,
+  "matched_mcc_abs_per_threshold": {
+   "0.6": 0.8593892807993077,
+   "0.7": 0.8901386476047501,
+   "0.8": 0.9187723668810305,
+   "0.9": 0.9497024896086197
+  },
+  "matched_mcc_abs_band": 0.904500696223427,
+  "act_iou_mean": 0.3752439797413662,
+  "act_iou_std": 0.0015233598583079414,
+  "act_iou_n_pairs": 12288.0,
+  "alive_source": "stats"
+ }
+}

lm/gpt2-small_layer8/batchtopk/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": 128,
+   "l0_raw": 127.7182478515625,
+   "recon_r2": 0.9548467507110822,
+   "l1_per_token": 607.7213934281276,
+   "stats_source": "stats",
+   "n_dead": 0,
+   "dead_rate": 0.0,
+   "alive_source": "stats",
+   "n_dense": 242,
+   "dense_rate": 0.019694010416666668,
+   "dense_threshold": 0.1,
+   "n_dup_signed": {
+    "0.5": 470,
+    "0.6": 95,
+    "0.7": 21,
+    "0.8": 5,
+    "0.9": 2
+   },
+   "n_dup_clusters_signed": {
+    "0.5": 232,
+    "0.6": 47,
+    "0.7": 10,
+    "0.8": 2,
+    "0.9": 1
+   },
+   "largest_dup_signed": {
+    "0.5": 3,
+    "0.6": 3,
+    "0.7": 3,
+    "0.8": 3,
+    "0.9": 2
+   },
+   "babel_mu1_signed": 0.3079283535480499,
+   "n_dup_abs": {
+    "0.5": 1342,
+    "0.6": 915,
+    "0.7": 807,
+    "0.8": 751,
+    "0.9": 656
+   },
+   "n_dup_clusters_abs": {
+    "0.5": 667,
+    "0.6": 456,
+    "0.7": 403,
+    "0.8": 375,
+    "0.9": 328
+   },
+   "largest_dup_abs": {
+    "0.5": 3,
+    "0.6": 3,
+    "0.7": 3,
+    "0.8": 3,
+    "0.9": 2
+   },
+   "babel_mu1_abs": 0.36509886384010315
+  },
+  "seed1": {
+   "d_sae": 12288,
+   "d_in": 768,
+   "l0": 128,
+   "l0_raw": 128.3524513671875,
+   "recon_r2": 0.9549618942987953,
+   "l1_per_token": 610.2288667752479,
+   "stats_source": "stats",
+   "n_dead": 0,
+   "dead_rate": 0.0,
+   "alive_source": "stats",
+   "n_dense": 238,
+   "dense_rate": 0.019368489583333332,
+   "dense_threshold": 0.1,
+   "n_dup_signed": {
+    "0.5": 497,
+    "0.6": 95,
+    "0.7": 19,
+    "0.8": 5,
+    "0.9": 3
+   },
+   "n_dup_clusters_signed": {
+    "0.5": 245,
+    "0.6": 46,
+    "0.7": 9,
+    "0.8": 2,
+    "0.9": 1
+   },
+   "largest_dup_signed": {
+    "0.5": 3,
+    "0.6": 3,
+    "0.7": 3,
+    "0.8": 3,
+    "0.9": 3
+   },
+   "babel_mu1_signed": 0.30881133675575256,
+   "n_dup_abs": {
+    "0.5": 1375,
+    "0.6": 928,
+    "0.7": 807,
+    "0.8": 739,
+    "0.9": 631
+   },
+   "n_dup_clusters_abs": {
+    "0.5": 681,
+    "0.6": 461,
+    "0.7": 402,
+    "0.8": 369,
+    "0.9": 315
+   },
+   "largest_dup_abs": {
+    "0.5": 3,
+    "0.6": 3,
+    "0.7": 3,
+    "0.8": 3,
+    "0.9": 3
+   },
+   "babel_mu1_abs": 0.3659749925136566
+  },
+  "seed2": {
+   "d_sae": 12288,
+   "d_in": 768,
+   "l0": 127,
+   "l0_raw": 127.48565991210937,
+   "recon_r2": 0.954758292979353,
+   "l1_per_token": 606.2793811777946,
+   "stats_source": "stats",
+   "n_dead": 0,
+   "dead_rate": 0.0,
+   "alive_source": "stats",
+   "n_dense": 238,
+   "dense_rate": 0.019368489583333332,
+   "dense_threshold": 0.1,
+   "n_dup_signed": {
+    "0.5": 532,
+    "0.6": 123,
+    "0.7": 31,
+    "0.8": 9,
+    "0.9": 2
+   },
+   "n_dup_clusters_signed": {
+    "0.5": 261,
+    "0.6": 61,
+    "0.7": 15,
+    "0.8": 4,
+    "0.9": 1
+   },
+   "largest_dup_signed": {
+    "0.5": 4,
+    "0.6": 3,
+    "0.7": 3,
+    "0.8": 3,
+    "0.9": 2
+   },
+   "babel_mu1_signed": 0.30918964743614197,
+   "n_dup_abs": {
+    "0.5": 1405,
+    "0.6": 943,
+    "0.7": 785,
+    "0.8": 713,
+    "0.9": 638
+   },
+   "n_dup_clusters_abs": {
+    "0.5": 694,
+    "0.6": 469,
+    "0.7": 390,
+    "0.8": 356,
+    "0.9": 319
+   },
+   "largest_dup_abs": {
+    "0.5": 4,
+    "0.6": 3,
+    "0.7": 3,
+    "0.8": 3,
+    "0.9": 2
+   },
+   "babel_mu1_abs": 0.36544737219810486
+  }
+ },
+ "per_pair": {
+  "seed0\u2194seed1": {
+   "dcs_signed": 0.7136623859405518,
+   "dcs_abs": 0.7151963114738464,
+   "frac_sign_flipped": 0.022786458333333332,
+   "alive_source": "stats",
+   "n_alive_a": 12288,
+   "n_alive_b": 12288,
+   "matched_cos_signed": {
+    "0.5": 0.7085920572280884,
+    "0.6": 0.7129131555557251,
+    "0.7": 0.7134965658187866,
+    "0.8": 0.7136114835739136,
+    "0.9": 0.7136274576187134
+   },
+   "unique_frac_signed": {
+    "0.5": 0.9800618489583334,
+    "0.6": 0.9960530598958334,
+    "0.7": 0.9991455078125,
+    "0.8": 0.999755859375,
+    "0.9": 0.9998779296875
+   },
+   "dedup_dcs_signed": {
+    "0.5": 0.6944640417641494,
+    "0.6": 0.7100993300312742,
+    "0.7": 0.7128868885774864,
+    "0.8": 0.7134372620203067,
+    "0.9": 0.7135403448919533
+   },
+   "matched_frac_signed": {
+    "0.5": 0.7732747395833334,
+    "0.6": 0.7207845052083334,
+    "0.7": 0.6289876302083334,
+    "0.8": 0.5049641927083334,
+    "0.9": 0.2537434895833333
+   },
+   "matched_cos_abs": {
+    "0.5": 0.7161705493927002,
+    "0.6": 0.7201769948005676,
+    "0.7": 0.7203535437583923,
+    "0.8": 0.7201673984527588,
+    "0.9": 0.7195829153060913
+   },
+   "unique_frac_abs": {
+    "0.5": 0.9442952473958334,
+    "0.6": 0.9623209635416666,
+    "0.7": 0.9670817057291666,
+    "0.8": 0.9696451822916666,
+    "0.9": 0.9737955729166666
+   },
+   "dedup_dcs_abs": {
+    "0.5": 0.6762764461163897,
+    "0.6": 0.693041419557024,
+    "0.7": 0.696640733825916,
+    "0.8": 0.6983068483532406,
+    "0.9": 0.7007266572715404
+   },
+   "matched_frac_abs": {
+    "0.5": 0.7564290364583334,
+    "0.6": 0.7080078125,
+    "0.7": 0.6189778645833334,
+    "0.8": 0.4969889322916667,
+    "0.9": 0.24943033854166666
+   }
+  },
+  "seed0\u2194seed2": {
+   "dcs_signed": 0.7091007828712463,
+   "dcs_abs": 0.7105391621589661,
+   "frac_sign_flipped": 0.025797526041666668,
+   "alive_source": "stats",
+   "n_alive_a": 12288,
+   "n_alive_b": 12288,
+   "matched_cos_signed": {
+    "0.5": 0.7051715850830078,
+    "0.6": 0.7086870670318604,
+    "0.7": 0.7091602683067322,
+    "0.8": 0.7090668082237244,
+    "0.9": 0.7090792059898376
+   },
+   "unique_frac_signed": {
+    "0.5": 0.9792887369791666,
+    "0.6": 0.9955240885416666,
+    "0.7": 0.9989013671875,
+    "0.8": 0.9996744791666666,
+    "0.9": 0.9999186197916666
+   },
+   "dedup_dcs_signed": {
+    "0.5": 0.6905665909095356,
+    "0.6": 0.7055150464681598,
+    "0.7": 0.7083811615666491,
+    "0.8": 0.7088359922054224,
+    "0.9": 0.7090215009763293
+   },
+   "matched_frac_signed": {
+    "0.5": 0.7677408854166666,
+    "0.6": 0.7119140625,
+    "0.7": 0.6239420572916666,
+    "0.8": 0.4934895833333333,
+    "0.9": 0.2490234375
+   },
+   "matched_cos_abs": {
+    "0.5": 0.7125113606452942,
+    "0.6": 0.7156597971916199,
+    "0.7": 0.7157593965530396,
+    "0.8": 0.7155851125717163,
+    "0.9": 0.715090811252594
+   },
+   "unique_frac_abs": {
+    "0.5": 0.943603515625,
+    "0.6": 0.9620361328125,
+    "0.7": 0.9674886067708334,
+    "0.8": 0.9701741536458334,
+    "0.9": 0.9736735026041666
+   },
+   "dedup_dcs_abs": {
+    "0.5": 0.6723282248276519,
+    "0.6": 0.688490583699604,
+    "0.7": 0.6924890613542327,
+    "0.8": 0.6942421809508232,
+    "0.9": 0.6962649748723682
+   },
+   "matched_frac_abs": {
+    "0.5": 0.750244140625,
+    "0.6": 0.6988118489583334,
+    "0.7": 0.614501953125,
+    "0.8": 0.486572265625,
+    "0.9": 0.2451171875
+   }
+  },
+  "seed1\u2194seed2": {
+   "dcs_signed": 0.7128297686576843,
+   "dcs_abs": 0.7144290804862976,
+   "frac_sign_flipped": 0.027262369791666668,
+   "alive_source": "stats",
+   "n_alive_a": 12288,
+   "n_alive_b": 12288,
+   "matched_cos_signed": {
+    "0.5": 0.7082353830337524,
+    "0.6": 0.7123907804489136,
+    "0.7": 0.7128813862800598,
+    "0.8": 0.7128597497940063,
+    "0.9": 0.7127940654754639
+   },
+   "unique_frac_signed": {
+    "0.5": 0.9787190755208334,
+    "0.6": 0.9954833984375,
+    "0.7": 0.9989420572916666,
+    "0.8": 0.9996744791666666,
+    "0.9": 0.9998779296875
+   },
+   "dedup_dcs_signed": {
+    "0.5": 0.6931634793339375,
+    "0.6": 0.7091731951368274,
+    "0.7": 0.7121271986155382,
+    "0.8": 0.7126276990942035,
+    "0.9": 0.7127070544811431
+   },
+   "matched_frac_signed": {
+    "0.5": 0.7713216145833334,
+    "0.6": 0.7215983072916666,
+    "0.7": 0.6263020833333334,
+    "0.8": 0.503173828125,
+    "0.9": 0.2534993489583333
+   },
+   "matched_cos_abs": {
+    "0.5": 0.7156670689582825,
+    "0.6": 0.719334065914154,
+    "0.7": 0.7195640206336975,
+    "0.8": 0.7192935943603516,
+    "0.9": 0.7184290885925293
+   },
+   "unique_frac_abs": {
+    "0.5": 0.9428304036458334,
+    "0.6": 0.9617106119791666,
+    "0.7": 0.9674479166666666,
+    "0.8": 0.9704182942708334,
+    "0.9": 0.9741617838541666
+   },
+   "dedup_dcs_abs": {
+    "0.5": 0.6747526715019679,
+    "0.6": 0.6917912047477632,
+    "0.7": 0.6961407126703609,
+    "0.8": 0.6980156629191091,
+    "0.9": 0.6998661625160214
+   },
+   "matched_frac_abs": {
+    "0.5": 0.7534993489583334,
+    "0.6": 0.70947265625,
+    "0.7": 0.616943359375,
+    "0.8": 0.4955240885416667,
+    "0.9": 0.24951171875
+   }
+  }
+ },
+ "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": 239,
+  "n_dense_std": 2.3094010767585034,
+  "l0_mean": 128,
+  "l0_std": 0.4486430162394102,
+  "recon_r2_mean": 0.95485564599641,
+  "recon_r2_std": null,
+  "l1_per_token_mean": 608.0765471270566,
+  "l1_per_token_std": 1.9985449960435793,
+  "babel_mu1_signed_mean": 0.3086431125799815,
+  "babel_mu1_signed_std": 0.00064725582115573,
+  "n_dup_signed_per_threshold": {
+   "0.5": 499.6666666666667,
+   "0.6": 104.33333333333333,
+   "0.7": 23.666666666666668,
+   "0.8": 6.333333333333333,
+   "0.9": 2.3333333333333335
+  },
+  "n_dup_signed_band": 34.16666666666667,
+  "n_dup_signed_band_std": 6.135620044733321,
+  "n_dup_clusters_signed_per_threshold": {
+   "0.5": 246.0,
+   "0.6": 51.333333333333336,
+   "0.7": 11.333333333333334,
+   "0.8": 2.6666666666666665,
+   "0.9": 1.0
+  },
+  "n_dup_clusters_signed_band": 16.583333333333336,
+  "n_dup_clusters_signed_band_std": 3.185252475602732,
+  "largest_dup_signed_per_threshold": {
+   "0.5": 3.3333333333333335,
+   "0.6": 3.0,
+   "0.7": 3.0,
+   "0.8": 3.0,
+   "0.9": 2.3333333333333335
+  },
+  "largest_dup_signed_band": 2.8333333333333335,
+  "largest_dup_signed_band_std": 0.14433756729740643,
+  "dcs_signed_mean": 0.7118643124898275,
+  "dcs_signed_std": 0.0024292251541414983,
+  "dedup_dcs_signed_per_threshold": {
+   "0.5": 0.6927313706692075,
+   "0.6": 0.7082625238787538,
+   "0.7": 0.7111317495865578,
+   "0.8": 0.7116336511066441,
+   "0.9": 0.7117563001164752
+  },
+  "dedup_dcs_signed_band": 0.7106960561721076,
+  "dedup_dcs_signed_band_std": 0.0024241539848505723,
+  "matched_cos_signed_per_threshold": {
+   "0.5": 0.7073330084482828,
+   "0.6": 0.7113303343454996,
+   "0.7": 0.7118460734685262,
+   "0.8": 0.7118460138638815,
+   "0.9": 0.7118335763613383
+  },
+  "matched_cos_signed_band": 0.7117139995098114,
+  "matched_cos_signed_band_std": 0.0023763298445997004,
+  "unique_frac_signed_per_threshold": {
+   "0.5": 0.9793565538194445,
+   "0.6": 0.9956868489583334,
+   "0.7": 0.9989963107638888,
+   "0.8": 0.9997016059027777,
+   "0.9": 0.9998914930555555
+  },
+  "unique_frac_signed_band": 0.9985690646701388,
+  "unique_frac_signed_band_std": 0.00012050615535432782,
+  "matched_frac_signed_per_threshold": {
+   "0.5": 0.7707790798611112,
+   "0.6": 0.7180989583333334,
+   "0.7": 0.6264105902777778,
+   "0.8": 0.5005425347222222,
+   "0.9": 0.2520887586805555
+  },
+  "matched_frac_signed_band": 0.5242852105034722,
+  "matched_frac_signed_band_std": 0.0040934190861738485,
+  "babel_mu1_abs_mean": 0.3655070761839549,
+  "babel_mu1_abs_std": 0.00044110518881602777,
+  "n_dup_abs_per_threshold": {
+   "0.5": 1374.0,
+   "0.6": 928.6666666666666,
+   "0.7": 799.6666666666666,
+   "0.8": 734.3333333333334,
+   "0.9": 641.6666666666666
+  },
+  "n_dup_abs_band": 776.0833333333333,
+  "n_dup_abs_band_std": 6.251666444503684,
+  "n_dup_clusters_abs_per_threshold": {
+   "0.5": 680.6666666666666,
+   "0.6": 462.0,
+   "0.7": 398.3333333333333,
+   "0.8": 366.6666666666667,
+   "0.9": 320.6666666666667
+  },
+  "n_dup_clusters_abs_band": 386.9166666666667,
+  "n_dup_clusters_abs_band_std": 3.50297492616395,
+  "largest_dup_abs_per_threshold": {
+   "0.5": 3.3333333333333335,
+   "0.6": 3.0,
+   "0.7": 3.0,
+   "0.8": 3.0,
+   "0.9": 2.3333333333333335
+  },
+  "largest_dup_abs_band": 2.8333333333333335,
+  "largest_dup_abs_band_std": 0.14433756729740643,
+  "dcs_abs_mean": 0.71338818470637,
+  "dcs_abs_std": 0.0024969697542773614,
+  "dedup_dcs_abs_per_threshold": {
+   "0.5": 0.6744524474820032,
+   "0.6": 0.6911077360014638,
+   "0.7": 0.6950901692835032,
+   "0.8": 0.6968548974077243,
+   "0.9": 0.6989525982199766
+  },
+  "dedup_dcs_abs_band": 0.695501350228167,
+  "dedup_dcs_abs_band_std": 0.0023060516767600663,
+  "matched_cos_abs_per_threshold": {
+   "0.5": 0.7147829929987589,
+   "0.6": 0.7183902859687805,
+   "0.7": 0.7185589869817098,
+   "0.8": 0.7183487017949423,
+   "0.9": 0.7177009383837382
+  },
+  "matched_cos_abs_band": 0.7182497282822927,
+  "matched_cos_abs_band_std": 0.002404664994481575,
+  "unique_frac_abs_per_threshold": {
+   "0.5": 0.943576388888889,
+   "0.6": 0.9620225694444443,
+   "0.7": 0.9673394097222222,
+   "0.8": 0.9700792100694445,
+   "0.9": 0.973876953125
+  },
+  "unique_frac_abs_band": 0.9683295355902778,
+  "unique_frac_abs_band_std": 0.0001125126140885089,
+  "matched_frac_abs_per_threshold": {
+   "0.5": 0.753390842013889,
+   "0.6": 0.7054307725694445,
+   "0.7": 0.6168077256944445,
+   "0.8": 0.4930284288194445,
+   "0.9": 0.24801974826388887
+  },
+  "matched_frac_abs_band": 0.5158216688368056,
+  "matched_frac_abs_band_std": 0.003965998166427926,
+  "alive_source": "stats"
+ }
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/config.json

@@ -0,0 +1,12 @@
+{
+    "d_in": 768,
+    "d_sae_factor": 16,
+    "architecture": "batchtopk",
+    "activation_name": "relu",
+    "device": "cpu",
+    "seed": 0,
+    "top_k": 128,
+    "top_k_inference": 0,
+    "class_type": "BatchTopKSAE",
+    "module_name": "consdl.models.batchtopk"
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4542156a37ee3ec99e7ba415d096952c436b8464c17fed4820dd34d93c322549
+size 75550084

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e903fd90d7211921b1f0515505b4937f0b2d8dd8b97c858206a70538901b7e5
+size 99648

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45e5038b4f48b41b4f34484b212a519f5c329d0966db5b0bfc51e8a2697abca7
+size 50471

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json

@@ -0,0 +1,14 @@
+{
+  "n_samples": 20000,
+  "n_positions": 20480000,
+  "l0_per_token_0": 127.7182478515625,
+  "l0_per_token_0.001": 127.7182478515625,
+  "l0_per_token_0.01": 127.7182478515625,
+  "l0_per_token_0.05": 127.7182478515625,
+  "l0_per_token_0.1": 127.7182478515625,
+  "l0_per_token_0.5": 127.7182478515625,
+  "l0_per_token_1.0": 127.7182478515625,
+  "l1_per_token": 607.7213934281276,
+  "r2_score": 0.9548467507110822,
+  "l2_loss": 1.058953591401838
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/max_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ea7efe657c3b33aa7815ec8560dcf7665dde20243847cb0588d0c262a5dec73
+size 50382

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/mean_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c339d6727326b9cbe3da1f658fed6393fe63b7604c0c5b5ca0728ca54588955
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/meta.json

@@ -0,0 +1,26 @@
+{
+  "dataset": "fineweb",
+  "split": "train",
+  "hf_dataset": "HuggingFaceFW/fineweb",
+  "hf_dataset_config": null,
+  "hf_dataset_text_field": "text",
+  "context_length": 1024,
+  "num_samples_requested": 20000,
+  "num_samples_actual": 20000,
+  "batch_seqs": 32,
+  "stats_top_k": 100,
+  "stream_seed": 0,
+  "sparsity_thresholds": [
+    0.001,
+    0.01,
+    0.05,
+    0.1,
+    0.5,
+    1.0
+  ],
+  "model_hf_name": "openai-community/gpt2",
+  "hook_target_name": "transformer.h.8",
+  "tokenizer_hf_name": "openai-community/gpt2",
+  "d_sae": 12288,
+  "d_in": 768
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed0/stats.fineweb.split_train.seed_no/sparsity_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4e75a4e30e0d7f1bc6b5b245362a9956cabe7ebddc5f93699cc42a61da926af
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/config.json

@@ -0,0 +1,12 @@
+{
+    "d_in": 768,
+    "d_sae_factor": 16,
+    "architecture": "batchtopk",
+    "activation_name": "relu",
+    "device": "cpu",
+    "seed": 1,
+    "top_k": 128,
+    "top_k_inference": 0,
+    "class_type": "BatchTopKSAE",
+    "module_name": "consdl.models.batchtopk"
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c808be8f317e843dfd2f0bcba803f04b3317529ddb4d8f03dc89c0445b3ba893
+size 75550084

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f729fb21572179aef4947eba759e49089330dc258139f1b0fb4f193c9db06960
+size 99648

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57b51963dd53853f0f9514151a761dffc57ee52c9858e28ae401913c0a435387
+size 50471

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json

@@ -0,0 +1,14 @@
+{
+  "n_samples": 20000,
+  "n_positions": 20480000,
+  "l0_per_token_0": 128.3524513671875,
+  "l0_per_token_0.001": 128.3524513671875,
+  "l0_per_token_0.01": 128.3524513671875,
+  "l0_per_token_0.05": 128.3524513671875,
+  "l0_per_token_0.1": 128.3524513671875,
+  "l0_per_token_0.5": 128.3524513671875,
+  "l0_per_token_1.0": 128.3524513671875,
+  "l1_per_token": 610.2288667752479,
+  "r2_score": 0.9549618942987953,
+  "l2_loss": 1.0562531940294237
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/max_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9edd29217b79ee3dee47e88e274db68ffb05cdc1dc6ea9d33894806ce7eead3d
+size 50382

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/mean_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fde69c672847047db110cadad3443396db29fcd9613828ff9f5bda51a2914b8
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/meta.json

@@ -0,0 +1,26 @@
+{
+  "dataset": "fineweb",
+  "split": "train",
+  "hf_dataset": "HuggingFaceFW/fineweb",
+  "hf_dataset_config": null,
+  "hf_dataset_text_field": "text",
+  "context_length": 1024,
+  "num_samples_requested": 20000,
+  "num_samples_actual": 20000,
+  "batch_seqs": 32,
+  "stats_top_k": 100,
+  "stream_seed": 0,
+  "sparsity_thresholds": [
+    0.001,
+    0.01,
+    0.05,
+    0.1,
+    0.5,
+    1.0
+  ],
+  "model_hf_name": "openai-community/gpt2",
+  "hook_target_name": "transformer.h.8",
+  "tokenizer_hf_name": "openai-community/gpt2",
+  "d_sae": 12288,
+  "d_in": 768
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed1/stats.fineweb.split_train.seed_no/sparsity_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f268691c6c89f18ce1e2c30a68f361e818c37455ff20a39ad770ec1eb161695f
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/config.json

@@ -0,0 +1,12 @@
+{
+    "d_in": 768,
+    "d_sae_factor": 16,
+    "architecture": "batchtopk",
+    "activation_name": "relu",
+    "device": "cpu",
+    "seed": 2,
+    "top_k": 128,
+    "top_k_inference": 0,
+    "class_type": "BatchTopKSAE",
+    "module_name": "consdl.models.batchtopk"
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81d2f343f0dbeda4e2ffa6d049714ebc2c2f45a8f8ecbffa58115ccb8c7f4dac
+size 75550084

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:367f01c34387df42f6df2e31a1ed498cf7936b9e4ea04da1e2cfc1f56fc4938f
+size 99648

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78e83f048ecef584c406203f717dd5d89ea771421359414662583767ec50c008
+size 50471

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/l0_per_token.json

@@ -0,0 +1,14 @@
+{
+  "n_samples": 20000,
+  "n_positions": 20480000,
+  "l0_per_token_0": 127.48565991210937,
+  "l0_per_token_0.001": 127.48565991210937,
+  "l0_per_token_0.01": 127.48565991210937,
+  "l0_per_token_0.05": 127.48565991210937,
+  "l0_per_token_0.1": 127.48565991210937,
+  "l0_per_token_0.5": 127.48565991210937,
+  "l0_per_token_1.0": 127.48565991210937,
+  "l1_per_token": 606.2793811777946,
+  "r2_score": 0.954758292979353,
+  "l2_loss": 1.061028140502446
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/max_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d02e06c8cfeed8e8ce03a03f474e41cd868e20c1b45dc50467599f5fc79fa9b8
+size 50382

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/mean_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cca20f1e17748b6887b333f51da83fe12658a498e6a1a0a41dc8234f91be04ac
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/meta.json

@@ -0,0 +1,26 @@
+{
+  "dataset": "fineweb",
+  "split": "train",
+  "hf_dataset": "HuggingFaceFW/fineweb",
+  "hf_dataset_config": null,
+  "hf_dataset_text_field": "text",
+  "context_length": 1024,
+  "num_samples_requested": 20000,
+  "num_samples_actual": 20000,
+  "batch_seqs": 32,
+  "stats_top_k": 100,
+  "stream_seed": 0,
+  "sparsity_thresholds": [
+    0.001,
+    0.01,
+    0.05,
+    0.1,
+    0.5,
+    1.0
+  ],
+  "model_hf_name": "openai-community/gpt2",
+  "hook_target_name": "transformer.h.8",
+  "tokenizer_hf_name": "openai-community/gpt2",
+  "d_sae": 12288,
+  "d_in": 768
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_128/seed2/stats.fineweb.split_train.seed_no/sparsity_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61af4facba39ec6d7e84c402db4b1cfc7248324f7ac99a2476e09b071008d3b8
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_32/code_metrics_step_final.json

@@ -0,0 +1,136 @@
+{
+ "schema": 1,
+ "step": "final",
+ "n_tokens": 600000,
+ "thresholds": [
+  0.5,
+  0.6,
+  0.7,
+  0.8,
+  0.9
+ ],
+ "seeds": [
+  "seed0",
+  "seed1",
+  "seed2"
+ ],
+ "per_seed": {
+  "seed0": {
+   "l0": 31.919266666666665,
+   "recon_r2": 0.9193748679709128,
+   "l1_per_token": 304.20424793829403
+  },
+  "seed1": {
+   "l0": 31.818753333333333,
+   "recon_r2": 0.9194302259933855,
+   "l1_per_token": 306.4840420251751
+  },
+  "seed2": {
+   "l0": 31.798905,
+   "recon_r2": 0.9193767591649853,
+   "l1_per_token": 305.6252251752166
+  }
+ },
+ "per_pair": {
+  "seed0\u2194seed1": {
+   "alive_source": "stats",
+   "mcc_signed": 0.7400087756930157,
+   "mcc_abs": 0.7400210476221692,
+   "act_iou": 0.43364045041840676,
+   "act_iou_n_pairs": 12288,
+   "input_r2": 0.9686586865788722,
+   "matched_mcc_signed": {
+    "0.6": 0.8444406047600785,
+    "0.7": 0.8733211255663039,
+    "0.8": 0.9007055349628019,
+    "0.9": 0.9319587076178429
+   },
+   "matched_mcc_abs": {
+    "0.6": 0.8424924662374571,
+    "0.7": 0.8729844102628117,
+    "0.8": 0.9003978364915014,
+    "0.9": 0.9317171967888204
+   }
+  },
+  "seed0\u2194seed2": {
+   "alive_source": "stats",
+   "mcc_signed": 0.744693272009517,
+   "mcc_abs": 0.74470590875556,
+   "act_iou": 0.43729593880994333,
+   "act_iou_n_pairs": 12288,
+   "input_r2": 0.9695608845428498,
+   "matched_mcc_signed": {
+    "0.6": 0.8443687195173901,
+    "0.7": 0.8749583129128411,
+    "0.8": 0.9012890242036748,
+    "0.9": 0.9319492239243443
+   },
+   "matched_mcc_abs": {
+    "0.6": 0.8433518751414391,
+    "0.7": 0.8743328909963229,
+    "0.8": 0.9008074760540641,
+    "0.9": 0.9315127755633315
+   }
+  },
+  "seed1\u2194seed2": {
+   "alive_source": "stats",
+   "mcc_signed": 0.742542751517254,
+   "mcc_abs": 0.742554242294822,
+   "act_iou": 0.4358845957507067,
+   "act_iou_n_pairs": 12288,
+   "input_r2": 0.9696325740204559,
+   "matched_mcc_signed": {
+    "0.6": 0.8449652732784033,
+    "0.7": 0.8733104171762524,
+    "0.8": 0.9015445092077589,
+    "0.9": 0.9318724590191894
+   },
+   "matched_mcc_abs": {
+    "0.6": 0.844412736067978,
+    "0.7": 0.8730420062902585,
+    "0.8": 0.9014987154182973,
+    "0.9": 0.9316214859387513
+   }
+  }
+ },
+ "summary": {
+  "n_seeds": 3,
+  "n_pairs": 3,
+  "report_band": [
+   0.6,
+   0.7,
+   0.8,
+   0.9
+  ],
+  "l0_mean": 31.845641666666666,
+  "l0_std": 0.06452882730049893,
+  "recon_r2_mean": 0.9193939510430945,
+  "recon_r2_std": 3.142925655682355e-05,
+  "l1_per_token_mean": 305.4378383795619,
+  "l1_per_token_std": 1.1513907364720959,
+  "input_r2_mean": 0.969284048380726,
+  "input_r2_std": 0.0005427641133247415,
+  "mcc_signed_mean": 0.7424149330732622,
+  "mcc_signed_std": 0.0023448623841916204,
+  "mcc_abs_mean": 0.7424270662241838,
+  "mcc_abs_std": 0.002345018395342776,
+  "matched_mcc_signed_per_threshold": {
+   "0.6": 0.8445915325186241,
+   "0.7": 0.8738632852184658,
+   "0.8": 0.9011796894580786,
+   "0.9": 0.9319267968537922
+  },
+  "matched_mcc_signed_band": 0.8878903260122402,
+  "matched_mcc_abs_per_threshold": {
+   "0.6": 0.8434190258156247,
+   "0.7": 0.8734531025164644,
+   "0.8": 0.9009013426546209,
+   "0.9": 0.9316171527636343
+  },
+  "matched_mcc_abs_band": 0.887347655937586,
+  "act_iou_mean": 0.43560699499301897,
+  "act_iou_std": 0.0018434873150272191,
+  "act_iou_n_pairs": 12288.0,
+  "alive_source": "stats"
+ }
+}

lm/gpt2-small_layer8/batchtopk/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": 32,
+   "l0_raw": 31.96863642578125,
+   "recon_r2": 0.9192533434187236,
+   "l1_per_token": 304.34399717607556,
+   "stats_source": "stats",
+   "n_dead": 0,
+   "dead_rate": 0.0,
+   "alive_source": "stats",
+   "n_dense": 20,
+   "dense_rate": 0.0016276041666666667,
+   "dense_threshold": 0.1,
+   "n_dup_signed": {
+    "0.5": 1951,
+    "0.6": 660,
+    "0.7": 227,
+    "0.8": 138,
+    "0.9": 120
+   },
+   "n_dup_clusters_signed": {
+    "0.5": 877,
+    "0.6": 260,
+    "0.7": 51,
+    "0.8": 8,
+    "0.9": 7
+   },
+   "largest_dup_signed": {
+    "0.5": 128,
+    "0.6": 127,
+    "0.7": 127,
+    "0.8": 99,
+    "0.9": 51
+   },
+   "babel_mu1_signed": 0.40909543633461,
+   "n_dup_abs": {
+    "0.5": 2123,
+    "0.6": 801,
+    "0.7": 340,
+    "0.8": 230,
+    "0.9": 180
+   },
+   "n_dup_clusters_abs": {
+    "0.5": 959,
+    "0.6": 329,
+    "0.7": 107,
+    "0.8": 54,
+    "0.9": 37
+   },
+   "largest_dup_abs": {
+    "0.5": 128,
+    "0.6": 127,
+    "0.7": 127,
+    "0.8": 99,
+    "0.9": 51
+   },
+   "babel_mu1_abs": 0.41840028762817383
+  },
+  "seed1": {
+   "d_sae": 12288,
+   "d_in": 768,
+   "l0": 32,
+   "l0_raw": 31.867031298828124,
+   "recon_r2": 0.9193055911896232,
+   "l1_per_token": 306.4091118377043,
+   "stats_source": "stats",
+   "n_dead": 0,
+   "dead_rate": 0.0,
+   "alive_source": "stats",
+   "n_dense": 22,
+   "dense_rate": 0.0017903645833333333,
+   "dense_threshold": 0.1,
+   "n_dup_signed": {
+    "0.5": 1930,
+    "0.6": 651,
+    "0.7": 240,
+    "0.8": 138,
+    "0.9": 122
+   },
+   "n_dup_clusters_signed": {
+    "0.5": 871,
+    "0.6": 260,
+    "0.7": 59,
+    "0.8": 11,
+    "0.9": 22
+   },
+   "largest_dup_signed": {
+    "0.5": 124,
+    "0.6": 124,
+    "0.7": 80,
+    "0.8": 44,
+    "0.9": 22
+   },
+   "babel_mu1_signed": 0.4086414575576782,
+   "n_dup_abs": {
+    "0.5": 2112,
+    "0.6": 796,
+    "0.7": 352,
+    "0.8": 232,
+    "0.9": 189
+   },
+   "n_dup_clusters_abs": {
+    "0.5": 957,
+    "0.6": 331,
+    "0.7": 114,
+    "0.8": 57,
+    "0.9": 55
+   },
+   "largest_dup_abs": {
+    "0.5": 124,
+    "0.6": 124,
+    "0.7": 80,
+    "0.8": 44,
+    "0.9": 22
+   },
+   "babel_mu1_abs": 0.4184608459472656
+  },
+  "seed2": {
+   "d_sae": 12288,
+   "d_in": 768,
+   "l0": 32,
+   "l0_raw": 31.848321533203126,
+   "recon_r2": 0.9192398937198489,
+   "l1_per_token": 305.4881751320176,
+   "stats_source": "stats",
+   "n_dead": 0,
+   "dead_rate": 0.0,
+   "alive_source": "stats",
+   "n_dense": 20,
+   "dense_rate": 0.0016276041666666667,
+   "dense_threshold": 0.1,
+   "n_dup_signed": {
+    "0.5": 1969,
+    "0.6": 635,
+    "0.7": 243,
+    "0.8": 145,
+    "0.9": 132
+   },
+   "n_dup_clusters_signed": {
+    "0.5": 882,
+    "0.6": 245,
+    "0.7": 55,
+    "0.8": 7,
+    "0.9": 9
+   },
+   "largest_dup_signed": {
+    "0.5": 134,
+    "0.6": 134,
+    "0.7": 133,
+    "0.8": 133,
+    "0.9": 63
+   },
+   "babel_mu1_signed": 0.40897050499916077,
+   "n_dup_abs": {
+    "0.5": 2138,
+    "0.6": 770,
+    "0.7": 364,
+    "0.8": 245,
+    "0.9": 202
+   },
+   "n_dup_clusters_abs": {
+    "0.5": 964,
+    "0.6": 312,
+    "0.7": 115,
+    "0.8": 57,
+    "0.9": 44
+   },
+   "largest_dup_abs": {
+    "0.5": 134,
+    "0.6": 134,
+    "0.7": 133,
+    "0.8": 133,
+    "0.9": 63
+   },
+   "babel_mu1_abs": 0.41863396763801575
+  }
+ },
+ "per_pair": {
+  "seed0\u2194seed1": {
+   "dcs_signed": 0.8013520240783691,
+   "dcs_abs": 0.8026490211486816,
+   "frac_sign_flipped": 0.014241536458333334,
+   "alive_source": "stats",
+   "n_alive_a": 12288,
+   "n_alive_b": 12288,
+   "matched_cos_signed": {
+    "0.5": 0.7888076901435852,
+    "0.6": 0.7969455718994141,
+    "0.7": 0.7999643683433533,
+    "0.8": 0.8003712296485901,
+    "0.9": 0.8005580902099609
+   },
+   "unique_frac_signed": {
+    "0.5": 0.9132080078125,
+    "0.6": 0.9678141276041666,
+    "0.7": 0.9854736328125,
+    "0.8": 0.9895426432291666,
+    "0.9": 0.9913330078125
+   },
+   "dedup_dcs_signed": {
+    "0.5": 0.7203454992632032,
+    "0.6": 0.771295183415835,
+    "0.7": 0.7883437921918812,
+    "0.8": 0.7920014621510442,
+    "0.9": 0.7936196594964713
+   },
+   "matched_frac_signed": {
+    "0.5": 0.7733561197916666,
+    "0.6": 0.7854817708333334,
+    "0.7": 0.7486165364583334,
+    "0.8": 0.6634928385416666,
+    "0.9": 0.5017903645833334
+   },
+   "matched_cos_abs": {
+    "0.5": 0.7907242774963379,
+    "0.6": 0.798725962638855,
+    "0.7": 0.8015782237052917,
+    "0.8": 0.8020060062408447,
+    "0.9": 0.8020365834236145
+   },
+   "unique_frac_abs": {
+    "0.5": 0.9056396484375,
+    "0.6": 0.9618733723958334,
+    "0.7": 0.9808349609375,
+    "0.8": 0.9857177734375,
+    "0.9": 0.9887288411458334
+   },
+   "dedup_dcs_abs": {
+    "0.5": 0.7161112566827796,
+    "0.6": 0.7682732353035439,
+    "0.7": 0.7862159457363305,
+    "0.8": 0.7905515747552272,
+    "0.9": 0.7929967016849939
+   },
+   "matched_frac_abs": {
+    "0.5": 0.7698567708333334,
+    "0.6": 0.783935546875,
+    "0.7": 0.746826171875,
+    "0.8": 0.662841796875,
+    "0.9": 0.5011393229166666
+   }
+  },
+  "seed0\u2194seed2": {
+   "dcs_signed": 0.8060774207115173,
+   "dcs_abs": 0.807426929473877,
+   "frac_sign_flipped": 0.014485677083333334,
+   "alive_source": "stats",
+   "n_alive_a": 12288,
+   "n_alive_b": 12288,
+   "matched_cos_signed": {
+    "0.5": 0.7918343544006348,
+    "0.6": 0.8003867268562317,
+    "0.7": 0.8041903376579285,
+    "0.8": 0.8048174381256104,
+    "0.9": 0.8050143718719482
+   },
+   "unique_frac_signed": {
+    "0.5": 0.9120686848958334,
+    "0.6": 0.9678548177083334,
+    "0.7": 0.9851888020833334,
+    "0.8": 0.9890950520833334,
+    "0.9": 0.9903971354166666
+   },
+   "dedup_dcs_signed": {
+    "0.5": 0.7222073182735282,
+    "0.6": 0.7746581496176077,
+    "0.7": 0.7922793154042059,
+    "0.8": 0.7960409458804255,
+    "0.9": 0.7972839278712247
+   },
+   "matched_frac_signed": {
+    "0.5": 0.7779134114583334,
+    "0.6": 0.7908528645833334,
+    "0.7": 0.7503255208333334,
+    "0.8": 0.6700846354166666,
+    "0.9": 0.5111490885416666
+   },
+   "matched_cos_abs": {
+    "0.5": 0.7933586835861206,
+    "0.6": 0.802190899848938,
+    "0.7": 0.8059757351875305,
+    "0.8": 0.806588351726532,
+    "0.9": 0.8066371083259583
+   },
+   "unique_frac_abs": {
+    "0.5": 0.9048665364583334,
+    "0.6": 0.962158203125,
+    "0.7": 0.9803873697916666,
+    "0.8": 0.9851888020833334,
+    "0.9": 0.9877522786458334
+   },
+   "dedup_dcs_abs": {
+    "0.5": 0.7178837241857158,
+    "0.6": 0.771834554761881,
+    "0.7": 0.7901684311364079,
+    "0.8": 0.7946418120118324,
+    "0.9": 0.7967576417892512
+   },
+   "matched_frac_abs": {
+    "0.5": 0.7736002604166666,
+    "0.6": 0.7881673177083334,
+    "0.7": 0.7491861979166666,
+    "0.8": 0.6689453125,
+    "0.9": 0.5105794270833334
+   }
+  },
+  "seed1\u2194seed2": {
+   "dcs_signed": 0.8028092384338379,
+   "dcs_abs": 0.803881824016571,
+   "frac_sign_flipped": 0.013427734375,
+   "alive_source": "stats",
+   "n_alive_a": 12288,
+   "n_alive_b": 12288,
+   "matched_cos_signed": {
+    "0.5": 0.7915000319480896,
+    "0.6": 0.7982300519943237,
+    "0.7": 0.8018153309822083,
+    "0.8": 0.8023896217346191,
+    "0.9": 0.8024964332580566
+   },
+   "unique_frac_signed": {
+    "0.5": 0.9126790364583334,
+    "0.6": 0.9682210286458334,
+    "0.7": 0.9849853515625,
+    "0.8": 0.9892171223958334,
+    "0.9": 0.9909261067708334
+   },
+   "dedup_dcs_signed": {
+    "0.5": 0.7223854865151225,
+    "0.6": 0.7728631220379611,
+    "0.7": 0.7897763556757127,
+    "0.8": 0.7937375526526012,
+    "0.9": 0.795214666305886
+   },
+   "matched_frac_signed": {
+    "0.5": 0.7759602864583334,
+    "0.6": 0.7887369791666666,
+    "0.7": 0.7491048177083334,
+    "0.8": 0.6639811197916666,
+    "0.9": 0.50634765625
+   },
+   "matched_cos_abs": {
+    "0.5": 0.7931916117668152,
+    "0.6": 0.8001823425292969,
+    "0.7": 0.8033261895179749,
+    "0.8": 0.803741455078125,
+    "0.9": 0.8037148118019104
+   },
+   "unique_frac_abs": {
+    "0.5": 0.9052327473958334,
+    "0.6": 0.9624430338541666,
+    "0.7": 0.9801839192708334,
+    "0.8": 0.9852294921875,
+    "0.9": 0.9881184895833334
+   },
+   "dedup_dcs_abs": {
+    "0.5": 0.7180230219310033,
+    "0.6": 0.7701299213804305,
+    "0.7": 0.7874074128946329,
+    "0.8": 0.7918697856366634,
+    "0.9": 0.7941654658934567
+   },
+   "matched_frac_abs": {
+    "0.5": 0.7728678385416666,
+    "0.6": 0.784912109375,
+    "0.7": 0.74755859375,
+    "0.8": 0.663818359375,
+    "0.9": 0.505859375
+   }
+  }
+ },
+ "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": 21,
+  "n_dense_std": 1.1547005383792515,
+  "l0_mean": 32,
+  "l0_std": 0.06474222029066243,
+  "recon_r2_mean": 0.9192662761093985,
+  "recon_r2_std": null,
+  "l1_per_token_mean": 305.4137613819325,
+  "l1_per_token_std": 1.034566429027056,
+  "babel_mu1_signed_mean": 0.40890246629714966,
+  "babel_mu1_signed_std": 0.00023451253954667774,
+  "n_dup_signed_per_threshold": {
+   "0.5": 1950.0,
+   "0.6": 648.6666666666666,
+   "0.7": 236.66666666666666,
+   "0.8": 140.33333333333334,
+   "0.9": 124.66666666666667
+  },
+  "n_dup_signed_band": 287.5833333333333,
+  "n_dup_signed_band_std": 1.2583057392117918,
+  "n_dup_clusters_signed_per_threshold": {
+   "0.5": 876.6666666666666,
+   "0.6": 255.0,
+   "0.7": 55.0,
+   "0.8": 8.666666666666666,
+   "0.9": 12.666666666666666
+  },
+  "n_dup_clusters_signed_band": 82.83333333333334,
+  "n_dup_clusters_signed_band_std": 4.645786621588784,
+  "largest_dup_signed_per_threshold": {
+   "0.5": 128.66666666666666,
+   "0.6": 128.33333333333334,
+   "0.7": 113.33333333333333,
+   "0.8": 92.0,
+   "0.9": 45.333333333333336
+  },
+  "largest_dup_signed_band": 94.75,
+  "largest_dup_signed_band_std": 24.724734578959588,
+  "dcs_signed_mean": 0.8034128944079081,
+  "dcs_signed_std": 0.002419843742148464,
+  "dedup_dcs_signed_per_threshold": {
+   "0.5": 0.721646101350618,
+   "0.6": 0.7729388183571345,
+   "0.7": 0.7901331544239333,
+   "0.8": 0.7939266535613569,
+   "0.9": 0.7953727512245273
+  },
+  "dedup_dcs_signed_band": 0.7880928443917381,
+  "dedup_dcs_signed_band_std": 0.001882862504375287,
+  "matched_cos_signed_per_threshold": {
+   "0.5": 0.7907140254974365,
+   "0.6": 0.7985207835833231,
+   "0.7": 0.80199001232783,
+   "0.8": 0.8025260965029398,
+   "0.9": 0.8026896317799886
+  },
+  "matched_cos_signed_band": 0.8014316310485204,
+  "matched_cos_signed_band_std": 0.0020783429698631793,
+  "unique_frac_signed_per_threshold": {
+   "0.5": 0.9126519097222223,
+   "0.6": 0.9679633246527778,
+   "0.7": 0.9852159288194445,
+   "0.8": 0.9892849392361112,
+   "0.9": 0.9908854166666666
+  },
+  "unique_frac_signed_band": 0.98333740234375,
+  "unique_frac_signed_band_std": 0.00020345052083331483,
+  "matched_frac_signed_per_threshold": {
+   "0.5": 0.7757432725694445,
+   "0.6": 0.7883572048611112,
+   "0.7": 0.7493489583333334,
+   "0.8": 0.6658528645833334,
+   "0.9": 0.5064290364583334
+  },
+  "matched_frac_signed_band": 0.6774970160590279,
+  "matched_frac_signed_band_std": 0.0029055935077566897,
+  "babel_mu1_abs_mean": 0.41849836707115173,
+  "babel_mu1_abs_std": 0.00012127432870639753,
+  "n_dup_abs_per_threshold": {
+   "0.5": 2124.3333333333335,
+   "0.6": 789.0,
+   "0.7": 352.0,
+   "0.8": 235.66666666666666,
+   "0.9": 190.33333333333334
+  },
+  "n_dup_abs_band": 391.75,
+  "n_dup_abs_band_std": 3.774917217635375,
+  "n_dup_clusters_abs_per_threshold": {
+   "0.5": 960.0,
+   "0.6": 324.0,
+   "0.7": 112.0,
+   "0.8": 56.0,
+   "0.9": 45.333333333333336
+  },
+  "n_dup_clusters_abs_band": 134.33333333333334,
+  "n_dup_clusters_abs_band_std": 4.259792639710686,
+  "largest_dup_abs_per_threshold": {
+   "0.5": 128.66666666666666,
+   "0.6": 128.33333333333334,
+   "0.7": 113.33333333333333,
+   "0.8": 92.0,
+   "0.9": 45.333333333333336
+  },
+  "largest_dup_abs_band": 94.75,
+  "largest_dup_abs_band_std": 24.724734578959588,
+  "dcs_abs_mean": 0.8046525915463766,
+  "dcs_abs_std": 0.0024804563954531683,
+  "dedup_dcs_abs_per_threshold": {
+   "0.5": 0.7173393342664997,
+   "0.6": 0.7700792371486185,
+   "0.7": 0.7879305965891238,
+   "0.8": 0.792354390801241,
+   "0.9": 0.7946399364559006
+  },
+  "dedup_dcs_abs_band": 0.786251040248721,
+  "dedup_dcs_abs_band_std": 0.0019454711076706708,
+  "matched_cos_abs_per_threshold": {
+   "0.5": 0.7924248576164246,
+   "0.6": 0.8003664016723633,
+   "0.7": 0.8036267161369324,
+   "0.8": 0.8041119376818339,
+   "0.9": 0.8041295011838278
+  },
+  "matched_cos_abs_band": 0.8030586391687393,
+  "matched_cos_abs_band_std": 0.0021483269535711945,
+  "unique_frac_abs_per_threshold": {
+   "0.5": 0.905246310763889,
+   "0.6": 0.962158203125,
+   "0.7": 0.98046875,
+   "0.8": 0.9853786892361112,
+   "0.9": 0.9881998697916666
+  },
+  "unique_frac_abs_band": 0.9790513780381944,
+  "unique_frac_abs_band_std": 0.00021442887035812565,
+  "matched_frac_abs_per_threshold": {
+   "0.5": 0.7721082899305555,
+   "0.6": 0.7856716579861112,
+   "0.7": 0.7478569878472222,
+   "0.8": 0.6652018229166666,
+   "0.9": 0.505859375
+  },
+  "matched_frac_abs_band": 0.6761474609375,
+  "matched_frac_abs_band_std": 0.002816963304139469,
+  "alive_source": "stats"
+ }
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/config.json

@@ -0,0 +1,12 @@
+{
+    "d_in": 768,
+    "d_sae_factor": 16,
+    "architecture": "batchtopk",
+    "activation_name": "relu",
+    "device": "cpu",
+    "seed": 0,
+    "top_k": 32,
+    "top_k_inference": 0,
+    "class_type": "BatchTopKSAE",
+    "module_name": "consdl.models.batchtopk"
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c6c2199c9edf2af46210a92a87cc0603b21a08c29b150f42e1fe73dbd376c48
+size 75550084

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2714be279f3b7a7edb76b1986f25f063d27243ad7ce03de6156cad09182cbce
+size 99648

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a10b442de0cdce6f5c0c62cda2d5f28e61d69107db1738b5d27495aebed49c7
+size 50471

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/l0_per_token.json

@@ -0,0 +1,14 @@
+{
+  "n_samples": 20000,
+  "n_positions": 20480000,
+  "l0_per_token_0": 31.96863642578125,
+  "l0_per_token_0.001": 31.96863642578125,
+  "l0_per_token_0.01": 31.96863642578125,
+  "l0_per_token_0.05": 31.96863642578125,
+  "l0_per_token_0.1": 31.96863642578125,
+  "l0_per_token_0.5": 31.96863642578125,
+  "l0_per_token_1.0": 31.96863642578125,
+  "l1_per_token": 304.34399717607556,
+  "r2_score": 0.9192533434187236,
+  "l2_loss": 1.8937056209024563
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/max_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8064e976734d923553dc88762712f13d7102d0158b12dee30854a7fe18a887c
+size 50382

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/mean_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98f7dfb8b1c6e7b94e18ebf154a86be17a59bcefc099499faf1d4c8bbf808277
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/meta.json

@@ -0,0 +1,26 @@
+{
+  "dataset": "fineweb",
+  "split": "train",
+  "hf_dataset": "HuggingFaceFW/fineweb",
+  "hf_dataset_config": null,
+  "hf_dataset_text_field": "text",
+  "context_length": 1024,
+  "num_samples_requested": 20000,
+  "num_samples_actual": 20000,
+  "batch_seqs": 32,
+  "stats_top_k": 100,
+  "stream_seed": 0,
+  "sparsity_thresholds": [
+    0.001,
+    0.01,
+    0.05,
+    0.1,
+    0.5,
+    1.0
+  ],
+  "model_hf_name": "openai-community/gpt2",
+  "hook_target_name": "transformer.h.8",
+  "tokenizer_hf_name": "openai-community/gpt2",
+  "d_sae": 12288,
+  "d_in": 768
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed0/stats.fineweb.split_train.seed_no/sparsity_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d48a6d8329da1058c73a7b0d533ba639f0e04862d2bd5217fb55ec4d605dbc05
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/config.json

@@ -0,0 +1,12 @@
+{
+    "d_in": 768,
+    "d_sae_factor": 16,
+    "architecture": "batchtopk",
+    "activation_name": "relu",
+    "device": "cpu",
+    "seed": 1,
+    "top_k": 32,
+    "top_k_inference": 0,
+    "class_type": "BatchTopKSAE",
+    "module_name": "consdl.models.batchtopk"
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f74337b486fc1617db11683aadad32dba731d172d1f30a1bd3637b28051a396b
+size 75550084

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/active_count_hist_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3d7831b6dea24ba39fd2f653f5f34031076d45bb738ef459882af4b2c7b2385
+size 99648

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/firing_rate_per_latent.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08bac7104ddae5fd076e1471370107d34af9bdeb2d4d88f3716bf71a8d35748b
+size 50471

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/l0_per_token.json

@@ -0,0 +1,14 @@
+{
+  "n_samples": 20000,
+  "n_positions": 20480000,
+  "l0_per_token_0": 31.867031298828124,
+  "l0_per_token_0.001": 31.867031298828124,
+  "l0_per_token_0.01": 31.867031298828124,
+  "l0_per_token_0.05": 31.867031298828124,
+  "l0_per_token_0.1": 31.867031298828124,
+  "l0_per_token_0.5": 31.867031298828124,
+  "l0_per_token_1.0": 31.867031298828124,
+  "l1_per_token": 306.4091118377043,
+  "r2_score": 0.9193055911896232,
+  "l2_loss": 1.892480283512387
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/max_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b26d1ce015f466b7e7b329191c8a45245802592d3642d22a21b2e784701d61b
+size 50382

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/mean_act_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:350318a196a0217fa49aea67182701743561dcb62cc32a91ab968e4bc0cdd72d
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/meta.json

@@ -0,0 +1,26 @@
+{
+  "dataset": "fineweb",
+  "split": "train",
+  "hf_dataset": "HuggingFaceFW/fineweb",
+  "hf_dataset_config": null,
+  "hf_dataset_text_field": "text",
+  "context_length": 1024,
+  "num_samples_requested": 20000,
+  "num_samples_actual": 20000,
+  "batch_seqs": 32,
+  "stats_top_k": 100,
+  "stream_seed": 0,
+  "sparsity_thresholds": [
+    0.001,
+    0.01,
+    0.05,
+    0.1,
+    0.5,
+    1.0
+  ],
+  "model_hf_name": "openai-community/gpt2",
+  "hook_target_name": "transformer.h.8",
+  "tokenizer_hf_name": "openai-community/gpt2",
+  "d_sae": 12288,
+  "d_in": 768
+}

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed1/stats.fineweb.split_train.seed_no/sparsity_per_token.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8931c98b91e66273ba09c5502ff526cb51526fa467619142f247d773eba40381
+size 50387

lm/gpt2-small_layer8/batchtopk/x16/l0_32/seed2/config.json

@@ -0,0 +1,12 @@
+{
+    "d_in": 768,
+    "d_sae_factor": 16,
+    "architecture": "batchtopk",
+    "activation_name": "relu",
+    "device": "cpu",
+    "seed": 2,
+    "top_k": 32,
+    "top_k_inference": 0,
+    "class_type": "BatchTopKSAE",
+    "module_name": "consdl.models.batchtopk"
+}