ce_loco LOCO run m3_nofe_India

exp26/ce_loco/runs/m3_nofe_India/calibrator.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56089458eafd50eb7a766f268418457668fb56b6b71e6d584dac6e78c7500930
+size 4544

exp26/ce_loco/runs/m3_nofe_India/features.json

@@ -0,0 +1,130 @@
+{
+ "features": [
+  "n_ratio",
+  "n_partial",
+  "n_tsort",
+  "n_tset",
+  "n_wratio",
+  "n_jw",
+  "n_lev",
+  "n_sp_ratio",
+  "n_alt_best",
+  "n_jac",
+  "n_idf_jac",
+  "n_max_idf_shared",
+  "n_shared",
+  "n_ntok_s1",
+  "n_ntok_pool",
+  "n_ntok_delta",
+  "n_legal3",
+  "n_first_eq",
+  "n_pfirst_in_s1",
+  "n_s1first_in_pool",
+  "a_jac",
+  "a_idf_jac",
+  "a_max_idf_shared",
+  "a_tset",
+  "a_stno3",
+  "a_num_jac",
+  "a_s1stno_in_pool",
+  "a_city3",
+  "a_loc_overlap",
+  "a_city_in_pool",
+  "a_state3",
+  "a_district3",
+  "a_pool_empty",
+  "a_ntok_s1",
+  "a_ntok_pool",
+  "a_ntok_delta",
+  "a_street_ratio",
+  "x_s1name_paddr",
+  "x_pname_s1addr",
+  "p_is_domain",
+  "p_has_indic",
+  "p_translit_unknown",
+  "s1_name_freq",
+  "s1_addr_freq",
+  "pool_name_freq",
+  "log_s1_name_freq",
+  "log_s1_addr_freq",
+  "log_pool_name_freq",
+  "sim_name",
+  "sim_nameaddr",
+  "sim_addr",
+  "sim_word",
+  "sim_bm25_word",
+  "sim_bm25_word_norm",
+  "sim_ce",
+  "sim_ce_rank",
+  "sim_ce_gap",
+  "from_reverse",
+  "is_s2",
+  "c_base",
+  "c_s1_rank",
+  "c_s1_gap",
+  "c_s1_margin",
+  "c_s1_n70",
+  "c_s1_n",
+  "c_p_n",
+  "c_p_n_rel",
+  "c_p_rank",
+  "c_p_gap",
+  "c_p_margin",
+  "country"
+ ],
+ "categorical": [
+  "country"
+ ],
+ "models": [
+  "model_fold0.txt",
+  "model_fold1.txt",
+  "model_fold2.txt",
+  "model_fold3.txt",
+  "model_fold4.txt"
+ ],
+ "calibrator": "calibrator.pkl",
+ "learner": "lgb",
+ "freq_rate": "1",
+ "decode": {
+  "min_prob": 0.0,
+  "max_k": null,
+  "threshold_oof": 0.7
+ },
+ "variant": "loco:india",
+ "holdout_country": "india",
+ "train_config": {
+  "cands": "/w/cands_m3-US.parquet",
+  "norm_dir": "/w/norm",
+  "pairs": "/w/exp26/ce_loco/data/pairs.parquet",
+  "gt": "/w/exp26/ce_loco/data/gt.parquet",
+  "out_dir": "/w/out/m3_nofe_India",
+  "s1_table": "s1",
+  "pool_tables": "s2,s3",
+  "max_s1": 300000,
+  "holdout_country": "India",
+  "folds": 5,
+  "fold_seed": 2026,
+  "folds_file": "/w/exp26/ce_loco/data/folds.parquet",
+  "pair_audit": null,
+  "eval_universe": "cands",
+  "learner": "lgb",
+  "exclude_s1": null,
+  "neg_ratio": 0.0,
+  "hard_neg_rank": 3,
+  "inner_val_frac": 0.1,
+  "drop_features": "sim_e5_full,sim_e5_name",
+  "refit": false,
+  "workers": 14,
+  "task_rows": 50000,
+  "lr": 0.05,
+  "leaves": 127,
+  "min_data_in_leaf": 200,
+  "feature_fraction": 0.8,
+  "rounds": 3000,
+  "early_stop": 100,
+  "log_period": 100,
+  "min_prob": 0.0,
+  "max_k": null,
+  "seed": 0
+ }
+}

exp26/ce_loco/runs/m3_nofe_India/folds.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab4f5bd6ae149ab9e661dd0c18b2638394a1fdb69a3e0014a2c2faf87a384467
+size 9856056

exp26/ce_loco/runs/m3_nofe_India/importance.csv

@@ -0,0 +1,72 @@
+feature,gain,split,gain_pct
+sim_ce,15426772.090962017,2156.2,65.62878591344987
+c_p_margin,4842141.220208031,1446.6,20.599503747767287
+sim_ce_gap,1287944.5896245688,1301.0,5.479191579577281
+c_p_rank,1266507.1562153487,144.2,5.387992155650366
+c_p_gap,227933.38396117688,168.4,0.9696773356287655
+sim_ce_rank,159918.08497498854,861.4,0.6803257156214281
+s1_name_freq,26348.82498857081,881.8,0.11209353350458659
+a_num_jac,24564.35178370476,284.8,0.10450200306387605
+sim_bm25_word_norm,17590.55444945544,1331.2,0.07483397857018596
+c_s1_rank,16679.21620386541,992.4,0.0709569508769067
+a_tset,15046.245172894001,738.4,0.06400994306719938
+sim_bm25_word,14942.128365030885,1541.4,0.06356700791181172
+pool_name_freq,12583.228652834707,1165.0,0.053531744326520275
+c_s1_margin,11635.998053187132,582.8,0.04950202288716875
+n_legal3,11077.679322147369,394.6,0.04712681566592869
+a_max_idf_shared,10593.544845592976,786.8,0.045067204119990115
+n_max_idf_shared,9312.647142001986,1699.0,0.03961799149985602
+s1_addr_freq,8931.109407901764,389.4,0.03799484842614482
+n_ntok_delta,8386.54109261036,240.8,0.035678138412619305
+c_base,8350.79408190772,549.0,0.03552606298824561
+n_tsort,7809.946716403962,354.6,0.03322518269046154
+c_s1_n70,7668.172428137064,1066.4,0.03262204456423285
+c_p_n,6438.863661652804,1044.2,0.027392302309572267
+log_pool_name_freq,6378.00418998003,247.4,0.027133393108498497
+a_idf_jac,6372.914128270745,497.4,0.02711173889799774
+log_s1_name_freq,5497.897287398577,174.4,0.023389230224008396
+x_pname_s1addr,5300.999211859703,1169.0,0.02255158372413681
+c_s1_n,4196.60585552454,718.6,0.017853258324643795
+c_s1_gap,4120.410268286057,603.6,0.01752910600989263
+n_wratio,3981.755512201786,273.4,0.01693923904036195
+n_idf_jac,3902.6141442418098,438.0,0.016602554744767427
+n_jw,3456.725737923384,405.0,0.01470565015662578
+x_s1name_paddr,3142.5271072387695,582.2,0.013368982022429358
+a_jac,2927.7510821342466,228.6,0.012455278903733226
+n_lev,2793.399498575926,356.0,0.01188371854992157
+n_partial,2354.477838349342,328.2,0.010016451988781228
+a_ntok_delta,2239.0955947875977,316.6,0.00952559126196958
+n_ratio,2085.0558746099473,243.4,0.008870273366683534
+n_tset,2060.953563052416,224.4,0.00876773698150188
+log_s1_addr_freq,1671.483533501625,65.4,0.007110848227433308
+a_stno3,1653.1350275754928,52.8,0.007032789761270883
+n_sp_ratio,1461.1348561048508,243.8,0.006215979992221741
+c_p_n_rel,1316.424960693717,199.2,0.005600353165721916
+n_ntok_pool,1256.025149345398,188.0,0.005343399457919755
+n_jac,1183.0841397285462,91.4,0.005033092812029796
+a_ntok_s1,1039.7030822753907,243.4,0.004423119146239391
+a_ntok_pool,1033.5098241567612,137.0,0.004396771702407483
+is_s2,960.8558679103851,201.6,0.004087686242912552
+n_ntok_s1,827.196944284439,173.4,0.0035190726125077522
+n_s1first_in_pool,787.6044977426529,40.8,0.0033506378821209003
+a_street_ratio,706.5867287397384,58.0,0.0030059709754131324
+n_shared,699.3844806492328,145.2,0.002975331072571483
+n_first_eq,579.4033101081848,44.8,0.0024649055273793822
+a_s1stno_in_pool,320.60140500068667,14.0,0.0013639069047158786
+n_pfirst_in_s1,274.3959023475647,9.8,0.0011673388201052515
+p_is_domain,144.25743794441223,8.6,0.0006137019757246067
+a_city_in_pool,94.41064851284027,18.2,0.00040164307884143845
+a_city3,56.48795399665833,11.2,0.00024031183047731867
+a_loc_overlap,52.09664626121521,9.0,0.00022163026873840398
+a_pool_empty,0.822398018836975,0.4,3.498656957894514e-06
+a_district3,0.0,0.0,0.0
+n_alt_best,0.0,0.0,0.0
+a_state3,0.0,0.0,0.0
+sim_word,0.0,0.0,0.0
+sim_name,0.0,0.0,0.0
+sim_nameaddr,0.0,0.0,0.0
+sim_addr,0.0,0.0,0.0
+p_translit_unknown,0.0,0.0,0.0
+p_has_indic,0.0,0.0,0.0
+from_reverse,0.0,0.0,0.0
+country,0.0,0.0,0.0

exp26/ce_loco/runs/m3_nofe_India/metrics.json

@@ -0,0 +1,668 @@
+{
+ "config": {
+  "cands": "/w/cands_m3-US.parquet",
+  "norm_dir": "/w/norm",
+  "pairs": "/w/exp26/ce_loco/data/pairs.parquet",
+  "gt": "/w/exp26/ce_loco/data/gt.parquet",
+  "out_dir": "/w/out/m3_nofe_India",
+  "s1_table": "s1",
+  "pool_tables": "s2,s3",
+  "max_s1": 300000,
+  "holdout_country": "India",
+  "folds": 5,
+  "fold_seed": 2026,
+  "folds_file": "/w/exp26/ce_loco/data/folds.parquet",
+  "pair_audit": null,
+  "eval_universe": "cands",
+  "learner": "lgb",
+  "exclude_s1": null,
+  "neg_ratio": 0.0,
+  "hard_neg_rank": 3,
+  "inner_val_frac": 0.1,
+  "drop_features": "sim_e5_full,sim_e5_name",
+  "refit": false,
+  "workers": 14,
+  "task_rows": 50000,
+  "lr": 0.05,
+  "leaves": 127,
+  "min_data_in_leaf": 200,
+  "feature_fraction": 0.8,
+  "rounds": 3000,
+  "early_stop": 100,
+  "log_period": 100,
+  "min_prob": 0.0,
+  "max_k": null,
+  "seed": 0
+ },
+ "universe_mode": "cands",
+ "n_eval_s1": 300000,
+ "eval_s1_by_country": {
+  "india": 120228,
+  "us": 179772
+ },
+ "n_rows": 10827771,
+ "n_pos": 1009021,
+ "folds": 5,
+ "fold_seed": 2026,
+ "folds_file": "/w/exp26/ce_loco/data/folds.parquet",
+ "results": {
+  "loco:india": {
+   "variant": "loco:india",
+   "protocol": "train ['us'] only (5 fold models averaged, isotonic fit on their OOF), score all 120,228 evaluated S1 ids of india (unseen category, nothing tuned on it)",
+   "n_s1": 120228,
+   "n_rows": 4463494,
+   "blocking_pair_recall": 0.962093672367323,
+   "pair_cal": {
+    "n": 4463494,
+    "pos": 401220,
+    "auc": 0.9993876359214834,
+    "logloss": 0.01903303057229191
+   },
+   "pair_raw": {
+    "n": 4463494,
+    "pos": 401220,
+    "auc": 0.9993614539914067,
+    "logloss": 0.01929144217283073
+   },
+   "macro_f05": 0.9556586453611203,
+   "n_s1_with_cands": 120228,
+   "expected_f05_mean_with_cands": 0.9849585909666713,
+   "realised_f05_mean_with_cands": 0.9556586453611202,
+   "expected_f05_mean_all": 0.9849585909666713,
+   "macro_f05_no_one_parent": 0.955528493883979,
+   "macro_f05_raw_probs": 0.9556713780578796,
+   "threshold_oof": 0.7,
+   "macro_f05_threshold_oof": 0.9553214367669947,
+   "per_source": {
+    "S2": {
+     "pairs_true": 201294,
+     "pairs_pred": 186307,
+     "tp": 183784,
+     "precision": 0.9864578357227587,
+     "recall": 0.9130128071378183,
+     "f05": 0.9708385013766188,
+     "blocking_recall": 0.9607986328454897
+    },
+    "S3": {
+     "pairs_true": 215734,
+     "pairs_pred": 197290,
+     "tp": 194615,
+     "precision": 0.9864412793349892,
+     "recall": 0.9021062975701558,
+     "f05": 0.9683359637931961,
+     "blocking_recall": 0.96330202935096
+    }
+   },
+   "noise": [],
+   "breakdown": [
+    {
+     "dimension": "overall",
+     "slice": "all",
+     "n": 120228,
+     "share": 1.0,
+     "macro_f05": 0.9556586453611203,
+     "mean_precision": 0.9844025449344598,
+     "mean_recall": 0.9079206017249525,
+     "n_singleton": 6729,
+     "singleton_acc": 0.9222767127359192,
+     "pred_empty_rate": 0.05982799347905646,
+     "mean_n_pred": 3.1905795654922313,
+     "mean_n_true": 3.4686429118009117,
+     "loss_share": 1.0
+    },
+    {
+     "dimension": "group",
+     "slice": "india",
+     "n": 120228,
+     "share": 1.0,
+     "macro_f05": 0.9556586453611203,
+     "mean_precision": 0.9844025449344598,
+     "mean_recall": 0.9079206017249525,
+     "n_singleton": 6729,
+     "singleton_acc": 0.9222767127359192,
+     "pred_empty_rate": 0.05982799347905646,
+     "mean_n_pred": 3.1905795654922313,
+     "mean_n_true": 3.4686429118009117,
+     "loss_share": 1.0
+    },
+    {
+     "dimension": "true_count",
+     "slice": "0",
+     "n": 6729,
+     "share": 0.05596865954686096,
+     "macro_f05": 0.9222767127359192,
+     "mean_precision": 0.0,
+     "mean_recall": null,
+     "n_singleton": 6729,
+     "singleton_acc": 0.9222767127359192,
+     "pred_empty_rate": 0.9222767127359192,
+     "mean_n_pred": 0.08114132857779759,
+     "mean_n_true": 0.0,
+     "loss_share": 0.09810408904224048
+    },
+    {
+     "dimension": "true_count",
+     "slice": "1",
+     "n": 6529,
+     "share": 0.054305153541604284,
+     "macro_f05": 0.8909454596389809,
+     "mean_precision": 0.9737931420260669,
+     "mean_recall": 0.9084086383826007,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.08730280287946086,
+     "mean_n_pred": 0.9523663654464696,
+     "mean_n_true": 1.0,
+     "loss_share": 0.13355982483948434
+    },
+    {
+     "dimension": "true_count",
+     "slice": "2",
+     "n": 20296,
+     "share": 0.1688125894134478,
+     "macro_f05": 0.947740115728882,
+     "mean_precision": 0.9836295852473204,
+     "mean_recall": 0.9087751281040599,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.01320457232952306,
+     "mean_n_pred": 1.8616968860859282,
+     "mean_n_true": 2.0,
+     "loss_share": 0.19895933396944682
+    },
+    {
+     "dimension": "true_count",
+     "slice": "3",
+     "n": 28752,
+     "share": 0.23914562331570016,
+     "macro_f05": 0.9609575827394998,
+     "mean_precision": 0.9888902463145892,
+     "mean_recall": 0.9100816175106659,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.0035823594880356148,
+     "mean_n_pred": 2.7711811352253757,
+     "mean_n_true": 3.0,
+     "loss_share": 0.21056693661152062
+    },
+    {
+     "dimension": "true_count",
+     "slice": "4",
+     "n": 26257,
+     "share": 0.2183933859001231,
+     "macro_f05": 0.9652885080985206,
+     "mean_precision": 0.9914290616166402,
+     "mean_recall": 0.9071771337167231,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.0010663822980538523,
+     "mean_n_pred": 3.668355105305252,
+     "mean_n_true": 4.0,
+     "loss_share": 0.1709636592690328
+    },
+    {
+     "dimension": "true_count",
+     "slice": "5+",
+     "n": 31665,
+     "share": 0.2633745882822637,
+     "macro_f05": 0.9683745000344709,
+     "mean_precision": 0.9932715731725117,
+     "mean_recall": 0.9059265342969748,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.0005684509711037423,
+     "mean_n_pred": 5.149249960524238,
+     "mean_n_true": 5.641023211747987,
+     "loss_share": 0.18784615626827503
+    }
+   ],
+   "threshold_sweep": {
+    "0.3": 0.952666888280845,
+    "0.35": 0.953636587328796,
+    "0.4": 0.9545037843489529,
+    "0.45": 0.9546956787029727,
+    "0.5": 0.9551594299349648,
+    "0.55": 0.9554636512173791,
+    "0.6": 0.9555040747345341,
+    "0.65": 0.9555191556300874,
+    "0.7": 0.9553214367669947,
+    "0.75": 0.9549827615532481,
+    "0.8": 0.9549760643212789,
+    "0.85": 0.9541828061049057,
+    "0.9": 0.9533415668708564
+   },
+   "threshold_sweep_best": 0.65
+  },
+  "loco:india/train_oof": {
+   "variant": "loco:india/train_oof",
+   "protocol": "OOF on the training countries ['us']",
+   "n_s1": 179772,
+   "n_rows": 6364277,
+   "blocking_pair_recall": 0.9761096041768044,
+   "pair_cal": {
+    "n": 6364277,
+    "pos": 607801,
+    "auc": 0.9999706710666612,
+    "logloss": 0.003801287747819355
+   },
+   "pair_raw": {
+    "n": 6364277,
+    "pos": 607801,
+    "auc": 0.9999703739113749,
+    "logloss": 0.0038279609061860043
+   },
+   "macro_f05": 0.986133913362099,
+   "n_s1_with_cands": 179772,
+   "expected_f05_mean_with_cands": 0.9932239048367307,
+   "realised_f05_mean_with_cands": 0.986133913362099,
+   "expected_f05_mean_all": 0.9932239048367307,
+   "macro_f05_no_one_parent": 0.9861217451712795,
+   "macro_f05_raw_probs": 0.9861178167119263,
+   "per_source": {
+    "S2": {
+     "pairs_true": 300881,
+     "pairs_pred": 287909,
+     "tp": 287476,
+     "precision": 0.9984960525721669,
+     "recall": 0.9554475025009888,
+     "f05": 0.989578779456626,
+     "blocking_recall": 0.9720288087316913
+    },
+    "S3": {
+     "pairs_true": 321796,
+     "pairs_pred": 310582,
+     "tp": 310165,
+     "precision": 0.9986573594091094,
+     "recall": 0.9638559832937638,
+     "f05": 0.991497477182116,
+     "blocking_recall": 0.9799251699834678
+    }
+   },
+   "noise": [],
+   "breakdown": [
+    {
+     "dimension": "overall",
+     "slice": "all",
+     "n": 179772,
+     "share": 1.0,
+     "macro_f05": 0.986133913362099,
+     "mean_precision": 0.9981598510730408,
+     "mean_recall": 0.9599410274492872,
+     "n_singleton": 9975,
+     "singleton_acc": 0.9889724310776943,
+     "pred_empty_rate": 0.05735042164519503,
+     "mean_n_pred": 3.3291669447967425,
+     "mean_n_true": 3.463704025098458,
+     "loss_share": 1.0
+    },
+    {
+     "dimension": "group",
+     "slice": "us",
+     "n": 179772,
+     "share": 1.0,
+     "macro_f05": 0.986133913362099,
+     "mean_precision": 0.9981598510730408,
+     "mean_recall": 0.9599410274492872,
+     "n_singleton": 9975,
+     "singleton_acc": 0.9889724310776943,
+     "pred_empty_rate": 0.05735042164519503,
+     "mean_n_pred": 3.3291669447967425,
+     "mean_n_true": 3.463704025098458,
+     "loss_share": 1.0
+    },
+    {
+     "dimension": "true_count",
+     "slice": "0",
+     "n": 9975,
+     "share": 0.05548695013684,
+     "macro_f05": 0.9889724310776943,
+     "mean_precision": 0.0,
+     "mean_recall": null,
+     "n_singleton": 9975,
+     "singleton_acc": 0.9889724310776943,
+     "pred_empty_rate": 0.9889724310776943,
+     "mean_n_pred": 0.011629072681704261,
+     "mean_n_true": 0.0,
+     "loss_share": 0.044128252109001016
+    },
+    {
+     "dimension": "true_count",
+     "slice": "1",
+     "n": 9814,
+     "share": 0.054591371292526086,
+     "macro_f05": 0.9573328874327448,
+     "mean_precision": 0.9961815867628341,
+     "mean_recall": 0.9600570613409415,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.03933156714897086,
+     "mean_n_pred": 0.966985938455268,
+     "mean_n_true": 1.0,
+     "loss_share": 0.1679822321153207
+    },
+    {
+     "dimension": "true_count",
+     "slice": "2",
+     "n": 30646,
+     "share": 0.1704714861046214,
+     "macro_f05": 0.9839739306717054,
+     "mean_precision": 0.9980332548814504,
+     "mean_recall": 0.9602721399203811,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.001762057038438948,
+     "mean_n_pred": 1.925993604385564,
+     "mean_n_true": 2.0,
+     "loss_share": 0.19702659633919978
+    },
+    {
+     "dimension": "true_count",
+     "slice": "3",
+     "n": 42971,
+     "share": 0.239030549807534,
+     "macro_f05": 0.9875763765901584,
+     "mean_precision": 0.9988797604363039,
+     "mean_recall": 0.960345349188988,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 9.308603476763399e-05,
+     "mean_n_pred": 2.885341276674967,
+     "mean_n_true": 3.0,
+     "loss_share": 0.21416464585899533
+    },
+    {
+     "dimension": "true_count",
+     "slice": "4",
+     "n": 39525,
+     "share": 0.21986182497830586,
+     "macro_f05": 0.9887111833980905,
+     "mean_precision": 0.9991486185608746,
+     "mean_recall": 0.960056925996205,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 2.5300442757748262e-05,
+     "mean_n_pred": 3.8443769765970903,
+     "mean_n_true": 4.0,
+     "loss_share": 0.1789964165633504
+    },
+    {
+     "dimension": "true_count",
+     "slice": "5+",
+     "n": 46841,
+     "share": 0.26055781768017267,
+     "macro_f05": 0.9894789144987511,
+     "mean_precision": 0.9994900766964772,
+     "mean_recall": 0.9592313707608001,
+     "n_singleton": 0,
+     "singleton_acc": null,
+     "pred_empty_rate": 0.0,
+     "mean_n_pred": 5.421020046540424,
+     "mean_n_true": 5.648000683162187,
+     "loss_share": 0.19770185701413265
+    }
+   ]
+  },
+  "loco:india/oracle_calibration": {
+   "variant": "loco:india/oracle_calibration",
+   "macro_f05": 0.9555459635860477,
+   "protocol": "DIAGNOSTIC ONLY: isotonic fit on india labels (not allowed for tuning)"
+  }
+ },
+ "best_iterations": [
+  217,
+  233,
+  225,
+  244,
+  228
+ ],
+ "headline": "loco:india",
+ "importance_top": [
+  {
+   "feature": "sim_ce",
+   "gain": 15426772.090962017,
+   "split": 2156.2,
+   "gain_pct": 65.62878591344987
+  },
+  {
+   "feature": "c_p_margin",
+   "gain": 4842141.220208031,
+   "split": 1446.6,
+   "gain_pct": 20.599503747767287
+  },
+  {
+   "feature": "sim_ce_gap",
+   "gain": 1287944.5896245688,
+   "split": 1301.0,
+   "gain_pct": 5.479191579577281
+  },
+  {
+   "feature": "c_p_rank",
+   "gain": 1266507.1562153487,
+   "split": 144.2,
+   "gain_pct": 5.387992155650366
+  },
+  {
+   "feature": "c_p_gap",
+   "gain": 227933.38396117688,
+   "split": 168.4,
+   "gain_pct": 0.9696773356287655
+  },
+  {
+   "feature": "sim_ce_rank",
+   "gain": 159918.08497498854,
+   "split": 861.4,
+   "gain_pct": 0.6803257156214281
+  },
+  {
+   "feature": "s1_name_freq",
+   "gain": 26348.82498857081,
+   "split": 881.8,
+   "gain_pct": 0.11209353350458659
+  },
+  {
+   "feature": "a_num_jac",
+   "gain": 24564.35178370476,
+   "split": 284.8,
+   "gain_pct": 0.10450200306387605
+  },
+  {
+   "feature": "sim_bm25_word_norm",
+   "gain": 17590.55444945544,
+   "split": 1331.2,
+   "gain_pct": 0.07483397857018596
+  },
+  {
+   "feature": "c_s1_rank",
+   "gain": 16679.21620386541,
+   "split": 992.4,
+   "gain_pct": 0.0709569508769067
+  },
+  {
+   "feature": "a_tset",
+   "gain": 15046.245172894001,
+   "split": 738.4,
+   "gain_pct": 0.06400994306719938
+  },
+  {
+   "feature": "sim_bm25_word",
+   "gain": 14942.128365030885,
+   "split": 1541.4,
+   "gain_pct": 0.06356700791181172
+  },
+  {
+   "feature": "pool_name_freq",
+   "gain": 12583.228652834707,
+   "split": 1165.0,
+   "gain_pct": 0.053531744326520275
+  },
+  {
+   "feature": "c_s1_margin",
+   "gain": 11635.998053187132,
+   "split": 582.8,
+   "gain_pct": 0.04950202288716875
+  },
+  {
+   "feature": "n_legal3",
+   "gain": 11077.679322147369,
+   "split": 394.6,
+   "gain_pct": 0.04712681566592869
+  },
+  {
+   "feature": "a_max_idf_shared",
+   "gain": 10593.544845592976,
+   "split": 786.8,
+   "gain_pct": 0.045067204119990115
+  },
+  {
+   "feature": "n_max_idf_shared",
+   "gain": 9312.647142001986,
+   "split": 1699.0,
+   "gain_pct": 0.03961799149985602
+  },
+  {
+   "feature": "s1_addr_freq",
+   "gain": 8931.109407901764,
+   "split": 389.4,
+   "gain_pct": 0.03799484842614482
+  },
+  {
+   "feature": "n_ntok_delta",
+   "gain": 8386.54109261036,
+   "split": 240.8,
+   "gain_pct": 0.035678138412619305
+  },
+  {
+   "feature": "c_base",
+   "gain": 8350.79408190772,
+   "split": 549.0,
+   "gain_pct": 0.03552606298824561
+  },
+  {
+   "feature": "n_tsort",
+   "gain": 7809.946716403962,
+   "split": 354.6,
+   "gain_pct": 0.03322518269046154
+  },
+  {
+   "feature": "c_s1_n70",
+   "gain": 7668.172428137064,
+   "split": 1066.4,
+   "gain_pct": 0.03262204456423285
+  },
+  {
+   "feature": "c_p_n",
+   "gain": 6438.863661652804,
+   "split": 1044.2,
+   "gain_pct": 0.027392302309572267
+  },
+  {
+   "feature": "log_pool_name_freq",
+   "gain": 6378.00418998003,
+   "split": 247.4,
+   "gain_pct": 0.027133393108498497
+  },
+  {
+   "feature": "a_idf_jac",
+   "gain": 6372.914128270745,
+   "split": 497.4,
+   "gain_pct": 0.02711173889799774
+  },
+  {
+   "feature": "log_s1_name_freq",
+   "gain": 5497.897287398577,
+   "split": 174.4,
+   "gain_pct": 0.023389230224008396
+  },
+  {
+   "feature": "x_pname_s1addr",
+   "gain": 5300.999211859703,
+   "split": 1169.0,
+   "gain_pct": 0.02255158372413681
+  },
+  {
+   "feature": "c_s1_n",
+   "gain": 4196.60585552454,
+   "split": 718.6,
+   "gain_pct": 0.017853258324643795
+  },
+  {
+   "feature": "c_s1_gap",
+   "gain": 4120.410268286057,
+   "split": 603.6,
+   "gain_pct": 0.01752910600989263
+  },
+  {
+   "feature": "n_wratio",
+   "gain": 3981.755512201786,
+   "split": 273.4,
+   "gain_pct": 0.01693923904036195
+  },
+  {
+   "feature": "n_idf_jac",
+   "gain": 3902.6141442418098,
+   "split": 438.0,
+   "gain_pct": 0.016602554744767427
+  },
+  {
+   "feature": "n_jw",
+   "gain": 3456.725737923384,
+   "split": 405.0,
+   "gain_pct": 0.01470565015662578
+  },
+  {
+   "feature": "x_s1name_paddr",
+   "gain": 3142.5271072387695,
+   "split": 582.2,
+   "gain_pct": 0.013368982022429358
+  },
+  {
+   "feature": "a_jac",
+   "gain": 2927.7510821342466,
+   "split": 228.6,
+   "gain_pct": 0.012455278903733226
+  },
+  {
+   "feature": "n_lev",
+   "gain": 2793.399498575926,
+   "split": 356.0,
+   "gain_pct": 0.01188371854992157
+  },
+  {
+   "feature": "n_partial",
+   "gain": 2354.477838349342,
+   "split": 328.2,
+   "gain_pct": 0.010016451988781228
+  },
+  {
+   "feature": "a_ntok_delta",
+   "gain": 2239.0955947875977,
+   "split": 316.6,
+   "gain_pct": 0.00952559126196958
+  },
+  {
+   "feature": "n_ratio",
+   "gain": 2085.0558746099473,
+   "split": 243.4,
+   "gain_pct": 0.008870273366683534
+  },
+  {
+   "feature": "n_tset",
+   "gain": 2060.953563052416,
+   "split": 224.4,
+   "gain_pct": 0.00876773698150188
+  },
+  {
+   "feature": "log_s1_addr_freq",
+   "gain": 1671.483533501625,
+   "split": 65.4,
+   "gain_pct": 0.007110848227433308
+  }
+ ],
+ "timings_s": {
+  "load": 6.9,
+  "features": 257.1,
+  "train": 252.8,
+  "eval": 66.9,
+  "total": 594.2
+ }
+}

exp26/ce_loco/runs/m3_nofe_India/oof.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb90fe2a2e3d0ce91095ecb349c5c3f8f61dd448879f1b9beece5c37d29092e2
+size 116821412

exp26/ce_loco/runs/m3_nofe_India/run.log

@@ -0,0 +1,143 @@
+[     3.5s rss=17.64GB] loaded: cands 79,652,854 rows, S1 2,206,821, pool 10,320,219, gt 2,206,821
+[     6.9s rss=17.09GB] folds: reused /w/exp26/ce_loco/data/folds.parquet (5 folds)
+[     9.2s rss=20.66GB] S1 universe: 2,206,821 ids (cands; candidates cover 100.0% of the 2,206,821 gt S1 ids in countries ['india', 'us'])
+[     9.3s rss=20.29GB] evaluated S1 ids: 300,000 {'india': 120228, 'us': 179772}
+[    32.9s rss=35.25GB] build_features: 79,652,854 candidate rows (10,827,771 featurised, 68,825,083 base-only), S1 table 2,206,821, pool table 10,320,219, workers 14
+[    42.7s rss=33.27GB]    packed tables: S1 456 MB, pool 2170 MB (9.8s)
+[    73.1s rss=35.58GB]    pair features: 8,000,000/79,652,854 rows, 266,260 rows/s (19,019/worker), eta 269s
+[   103.2s rss=36.47GB]    pair features: 23,977,771/79,652,854 rows, 398,506 rows/s (28,465/worker), eta 140s
+[   133.3s rss=36.47GB]    pair features: 43,027,771/79,652,854 rows, 477,028 rows/s (34,073/worker), eta 77s
+[   163.5s rss=36.47GB]    pair features: 65,427,771/79,652,854 rows, 543,020 rows/s (38,787/worker), eta 26s
+[   182.1s rss=36.47GB]    pair features: 79,652,854/79,652,854 rows, 572,763 rows/s (40,912/worker), eta 0s
+[   182.9s rss=32.57GB]    pair pass: 79,652,854 rows in 139.8s (570,509 rows/s overall)
+[   266.3s rss=36.81GB]    features assembled: 10,827,771 rows x 73 features (3.12 GB float32), parent-side 83.4s, total 257.0s
+[   269.3s rss=14.12GB] labels: 1,009,021 positive of 10,827,771 rows (0.0932); gt pairs of evaluated S1: 1,039,705
+[   272.6s rss=14.20GB] fold 1/5: fit 4,583,055 rows, early-stop 512,340, predict 1,268,882
+[100]	val's binary_logloss: 0.00521431
+[200]	val's binary_logloss: 0.00404559
+[300]	val's binary_logloss: 0.00404464
+[   306.0s rss=16.08GB]    trained on 4,583,055 rows: best_iter=217, val logloss=0.00404 (33s)
+[predict] 1,268,882/1,268,882 rows scored, 1,114,056 rows/s, eta 0s
+[predict] 2,000,000/4,463,494 rows scored, 952,099 rows/s, eta 3s
+[predict] 4,000,000/4,463,494 rows scored, 942,686 rows/s, eta 0s
+[predict] 4,463,494/4,463,494 rows scored, 954,439 rows/s, eta 0s
+[   312.2s rss=14.44GB] fold 2/5: fit 4,583,810 rows, early-stop 510,009, predict 1,270,458
+[100]	val's binary_logloss: 0.00506146
+[200]	val's binary_logloss: 0.0039062
+[300]	val's binary_logloss: 0.00391141
+[   357.5s rss=16.16GB]    trained on 4,583,810 rows: best_iter=233, val logloss=0.00389 (45s)
+[predict] 1,270,458/1,270,458 rows scored, 1,017,317 rows/s, eta 0s
+[predict] 2,000,000/4,463,494 rows scored, 845,454 rows/s, eta 3s
+[predict] 4,000,000/4,463,494 rows scored, 842,987 rows/s, eta 1s
+[predict] 4,463,494/4,463,494 rows scored, 853,702 rows/s, eta 0s
+[   364.4s rss=14.60GB] fold 3/5: fit 4,580,566 rows, early-stop 510,883, predict 1,272,828
+[100]	val's binary_logloss: 0.00491858
+[200]	val's binary_logloss: 0.00373944
+[300]	val's binary_logloss: 0.00376459
+[   409.3s rss=16.17GB]    trained on 4,580,566 rows: best_iter=225, val logloss=0.00374 (45s)
+[predict] 1,272,828/1,272,828 rows scored, 1,109,971 rows/s, eta 0s
+[predict] 2,000,000/4,463,494 rows scored, 886,494 rows/s, eta 3s
+[predict] 4,000,000/4,463,494 rows scored, 894,149 rows/s, eta 1s
+[predict] 4,463,494/4,463,494 rows scored, 904,103 rows/s, eta 0s
+[   415.8s rss=14.63GB] fold 4/5: fit 4,580,010 rows, early-stop 510,043, predict 1,274,224
+[100]	val's binary_logloss: 0.00518486
+[200]	val's binary_logloss: 0.0040371
+[300]	val's binary_logloss: 0.00402947
+[   463.6s rss=16.27GB]    trained on 4,580,010 rows: best_iter=244, val logloss=0.00402 (48s)
+[predict] 1,274,224/1,274,224 rows scored, 939,133 rows/s, eta 0s
+[predict] 2,000,000/4,463,494 rows scored, 777,864 rows/s, eta 3s
+[predict] 4,000,000/4,463,494 rows scored, 748,325 rows/s, eta 1s
+[predict] 4,463,494/4,463,494 rows scored, 758,260 rows/s, eta 0s
+[   471.3s rss=14.65GB] fold 5/5: fit 4,578,171 rows, early-stop 508,221, predict 1,277,885
+[100]	val's binary_logloss: 0.00498991
+[200]	val's binary_logloss: 0.00381944
+[300]	val's binary_logloss: 0.00380665
+[   517.5s rss=16.22GB]    trained on 4,578,171 rows: best_iter=228, val logloss=0.00380 (46s)
+[predict] 1,277,885/1,277,885 rows scored, 1,080,074 rows/s, eta 0s
+[predict] 2,000,000/4,463,494 rows scored, 909,591 rows/s, eta 3s
+[predict] 4,000,000/4,463,494 rows scored, 918,218 rows/s, eta 1s
+[predict] 4,463,494/4,463,494 rows scored, 927,756 rows/s, eta 0s
+
+==== [loco:india] macro F0.5 = 0.95566   (train ['us'] only (5 fold models averaged, isotonic fit on their OOF), score all 120,228 evaluated S1 ids of india (unseen category, nothing tuned on it))
+  S1 ids 120,228 (120,228 with candidates), rows 4,463,494, blocking pair recall 0.9621  [24s]
+  decoder expected F0.5 0.98496 vs realised 0.95566 (all S1); with candidates: expected 0.98496 vs realised 0.95566
+  pair-level diagnostics: calibrated AUC 0.99939 logloss 0.01903 | raw AUC 0.99936 logloss 0.01929
+  [loco:india/no_one_parent] 0.95553   [loco:india/raw_probs] 0.95567   [loco:india/threshold_oof t=0.7] 0.95532
+  [loco:india/sweep] plain threshold p > t (one-parent), diagnostic: 0.3:0.9527  0.35:0.9536  0.4:0.9545  0.45:0.9547  0.5:0.9552  0.55:0.9555  0.6:0.9555  0.65:0.9555  0.7:0.9553  0.75:0.9550  0.8:0.9550  0.85:0.9542  0.9:0.9533
+  source S2: pair P 0.9865 R 0.9130 F0.5 0.9708 (true 201,294, pred 186,307, blocking recall 0.9608)
+  source S3: pair P 0.9864 R 0.9021 F0.5 0.9683 (true 215,734, pred 197,290, blocking recall 0.9633)
+                       n  share  macro_f05  mean_precision  mean_recall  n_singleton  singleton_acc  pred_empty_rate  mean_n_pred  mean_n_true  loss_share
+dimension  slice                                                                                                                                          
+overall    all    120228 1.0000     0.9557          0.9844       0.9079         6729         0.9223           0.0598       3.1906       3.4686      1.0000
+group      india  120228 1.0000     0.9557          0.9844       0.9079         6729         0.9223           0.0598       3.1906       3.4686      1.0000
+true_count 0        6729 0.0560     0.9223          0.0000          NaN         6729         0.9223           0.9223       0.0811       0.0000      0.0981
+           1        6529 0.0543     0.8909          0.9738       0.9084            0            NaN           0.0873       0.9524       1.0000      0.1336
+           2       20296 0.1688     0.9477          0.9836       0.9088            0            NaN           0.0132       1.8617       2.0000      0.1990
+           3       28752 0.2391     0.9610          0.9889       0.9101            0            NaN           0.0036       2.7712       3.0000      0.2106
+           4       26257 0.2184     0.9653          0.9914       0.9072            0            NaN           0.0011       3.6684       4.0000      0.1710
+           5+      31665 0.2634     0.9684          0.9933       0.9059            0            NaN           0.0006       5.1492       5.6410      0.1878
+
+==== [loco:india/train_oof] macro F0.5 = 0.98613   (OOF on the training countries ['us'])
+  S1 ids 179,772 (179,772 with candidates), rows 6,364,277, blocking pair recall 0.9761  [14s]
+  decoder expected F0.5 0.99322 vs realised 0.98613 (all S1); with candidates: expected 0.99322 vs realised 0.98613
+  pair-level diagnostics: calibrated AUC 0.99997 logloss 0.00380 | raw AUC 0.99997 logloss 0.00383
+  [loco:india/train_oof/no_one_parent] 0.98612   [loco:india/train_oof/raw_probs] 0.98612
+  source S2: pair P 0.9985 R 0.9554 F0.5 0.9896 (true 300,881, pred 287,909, blocking recall 0.9720)
+  source S3: pair P 0.9987 R 0.9639 F0.5 0.9915 (true 321,796, pred 310,582, blocking recall 0.9799)
+                       n  share  macro_f05  mean_precision  mean_recall  n_singleton  singleton_acc  pred_empty_rate  mean_n_pred  mean_n_true  loss_share
+dimension  slice                                                                                                                                          
+overall    all    179772 1.0000     0.9861          0.9982       0.9599         9975         0.9890           0.0574       3.3292       3.4637      1.0000
+group      us     179772 1.0000     0.9861          0.9982       0.9599         9975         0.9890           0.0574       3.3292       3.4637      1.0000
+true_count 0        9975 0.0555     0.9890          0.0000          NaN         9975         0.9890           0.9890       0.0116       0.0000      0.0441
+           1        9814 0.0546     0.9573          0.9962       0.9601            0            NaN           0.0393       0.9670       1.0000      0.1680
+           2       30646 0.1705     0.9840          0.9980       0.9603            0            NaN           0.0018       1.9260       2.0000      0.1970
+           3       42971 0.2390     0.9876          0.9989       0.9603            0            NaN           0.0001       2.8853       3.0000      0.2142
+           4       39525 0.2199     0.9887          0.9991       0.9601            0            NaN           0.0000       3.8444       4.0000      0.1790
+           5+      46841 0.2606     0.9895          0.9995       0.9592            0            NaN           0.0000       5.4210       5.6480      0.1977
+
+  [loco:india/oracle_calibration] 0.95555  -> calibration transfer loss -0.00011 (diagnostic, uses india labels)
+
+feature importance (mean gain % over fold models, top 40):
+  sim_ce                  65.63%  splits     2156
+  c_p_margin              20.60%  splits     1447
+  sim_ce_gap               5.48%  splits     1301
+  c_p_rank                 5.39%  splits      144
+  c_p_gap                  0.97%  splits      168
+  sim_ce_rank              0.68%  splits      861
+  s1_name_freq             0.11%  splits      882
+  a_num_jac                0.10%  splits      285
+  sim_bm25_word_norm       0.07%  splits     1331
+  c_s1_rank                0.07%  splits      992
+  a_tset                   0.06%  splits      738
+  sim_bm25_word            0.06%  splits     1541
+  pool_name_freq           0.05%  splits     1165
+  c_s1_margin              0.05%  splits      583
+  n_legal3                 0.05%  splits      395
+  a_max_idf_shared         0.05%  splits      787
+  n_max_idf_shared         0.04%  splits     1699
+  s1_addr_freq             0.04%  splits      389
+  n_ntok_delta             0.04%  splits      241
+  c_base                   0.04%  splits      549
+  n_tsort                  0.03%  splits      355
+  c_s1_n70                 0.03%  splits     1066
+  c_p_n                    0.03%  splits     1044
+  log_pool_name_freq       0.03%  splits      247
+  a_idf_jac                0.03%  splits      497
+  log_s1_name_freq         0.02%  splits      174
+  x_pname_s1addr           0.02%  splits     1169
+  c_s1_n                   0.02%  splits      719
+  c_s1_gap                 0.02%  splits      604
+  n_wratio                 0.02%  splits      273
+  n_idf_jac                0.02%  splits      438
+  n_jw                     0.01%  splits      405
+  x_s1name_paddr           0.01%  splits      582
+  a_jac                    0.01%  splits      229
+  n_lev                    0.01%  splits      356
+  n_partial                0.01%  splits      328
+  a_ntok_delta             0.01%  splits      317
+  n_ratio                  0.01%  splits      243
+  n_tset                   0.01%  splits      224
+  log_s1_addr_freq         0.01%  splits       65
+
+HEADLINE [loco:india] macro F0.5 = 0.95566 (decoder expected 0.98496) over 120,228 S1 ids
+[   594.2s rss=10.60GB] saved ['model_fold0.txt', 'model_fold1.txt', 'model_fold2.txt', 'model_fold3.txt', 'model_fold4.txt'], calibrator.pkl, folds.parquet, features.json, metrics.json, oof.parquet in /w/out/m3_nofe_India; timings {'load': 6.9, 'features': 257.1, 'train': 252.8, 'eval': 66.9, 'total': 594.2}