Update multitask INDRA statement classifier

all_results.json

@@ -1,22 +1,28 @@
 {
     "epoch": 10.0,
-    "eval_gate1_accuracy": 0.9204524721766101,
-    "eval_gate1_f1": 0.9175696677724912,
-    "eval_gate1_precision": 0.9198718411071631,
-    "eval_gate1_recall": 0.9156173695177233,
+    "eval_gate1_accuracy": 0.9250136836343733,
+    "eval_gate1_f1": 0.9224529503160914,
+    "eval_gate1_precision": 0.9237612665493629,
+    "eval_gate1_recall": 0.9212694184275014,
     "eval_gate1_support": 5481,
-    "eval_gate2_accuracy": 0.8573253056011677,
-    "eval_gate2_f1": 0.7662196765345953,
-    "eval_gate2_precision": 0.7540255243234163,
-    "eval_gate2_recall": 0.7899899686908007,
+    "eval_gate2_accuracy": 0.8680897646414888,
+    "eval_gate2_f1": 0.8167601906166684,
+    "eval_gate2_precision": 0.8143474395151585,
+    "eval_gate2_recall": 0.8232715925733067,
     "eval_gate2_support": 5481,
-    "eval_gate3_accuracy": 0.8761520737327189,
-    "eval_gate3_f1": 0.7204269422018509,
-    "eval_gate3_precision": 0.7065840948237819,
-    "eval_gate3_recall": 0.7411799394675562,
+    "eval_gate3_accuracy": 0.8945852534562212,
+    "eval_gate3_f1": 0.814049046653749,
+    "eval_gate3_precision": 0.8216561919229995,
+    "eval_gate3_recall": 0.8169058833981815,
     "eval_gate3_support": 6944,
-    "eval_loss": 0.7167699933052063,
-    "eval_runtime": 9.951,
-    "eval_samples_per_second": 1248.614,
-    "eval_steps_per_second": 39.091
+    "eval_gate4_span_FN": 200,
+    "eval_gate4_span_FP": 210,
+    "eval_gate4_span_TP": 811,
+    "eval_gate4_span_f1": 0.7982283414489575,
+    "eval_gate4_span_precision": 0.794319294801231,
+    "eval_gate4_span_recall": 0.8021760632957252,
+    "eval_loss": NaN,
+    "eval_runtime": 12.712,
+    "eval_samples_per_second": 1057.505,
+    "eval_steps_per_second": 33.118
 }

config.json

@@ -6,10 +6,12 @@
   "classifier_dropout": null,
   "dtype": "float32",
   "gate1_loss_weight": 1.0,
-  "gate2_loss_weight": 0.5,
+  "gate2_loss_weight": 1.0,
   "gate2_num_labels": 20,
-  "gate3_loss_weight": 0.25,
+  "gate3_loss_weight": 1.0,
   "gate3_num_labels": 26,
+  "gate4_loss_weight": 1.0,
+  "gate4_num_labels": 3,
   "hidden_act": "gelu",
   "hidden_dropout_prob": 0.1,
   "hidden_size": 768,

eval_results.json

@@ -1,22 +1,28 @@
 {
     "epoch": 10.0,
-    "eval_gate1_accuracy": 0.9231751824817518,
-    "eval_gate1_f1": 0.91874922234221,
-    "eval_gate1_precision": 0.9213536363863121,
-    "eval_gate1_recall": 0.916446997689221,
+    "eval_gate1_accuracy": 0.9268248175182482,
+    "eval_gate1_f1": 0.922882664177459,
+    "eval_gate1_precision": 0.9238243397462302,
+    "eval_gate1_recall": 0.9219854713816811,
     "eval_gate1_support": 5480,
-    "eval_gate2_accuracy": 0.8593065693430657,
-    "eval_gate2_f1": 0.7763067392280264,
-    "eval_gate2_precision": 0.7921477517391428,
-    "eval_gate2_recall": 0.775442797363134,
+    "eval_gate2_accuracy": 0.8728102189781022,
+    "eval_gate2_f1": 0.7972864171672726,
+    "eval_gate2_precision": 0.8049759563676634,
+    "eval_gate2_recall": 0.7924699515870988,
     "eval_gate2_support": 5480,
-    "eval_gate3_accuracy": 0.8810483870967742,
-    "eval_gate3_f1": 0.778218086166726,
-    "eval_gate3_precision": 0.75635112362015,
-    "eval_gate3_recall": 0.8147429373450497,
+    "eval_gate3_accuracy": 0.8955933179723502,
+    "eval_gate3_f1": 0.7963151574821563,
+    "eval_gate3_precision": 0.7875457145835848,
+    "eval_gate3_recall": 0.8152968118858834,
     "eval_gate3_support": 6944,
-    "eval_loss": 0.7152989506721497,
-    "eval_runtime": 9.7809,
-    "eval_samples_per_second": 1270.237,
-    "eval_steps_per_second": 39.772
+    "eval_gate4_span_FN": 196,
+    "eval_gate4_span_FP": 216,
+    "eval_gate4_span_TP": 811,
+    "eval_gate4_span_f1": 0.7974434561529178,
+    "eval_gate4_span_precision": 0.7896786757469361,
+    "eval_gate4_span_recall": 0.8053624627526776,
+    "eval_loss": NaN,
+    "eval_runtime": 12.3946,
+    "eval_samples_per_second": 1084.506,
+    "eval_steps_per_second": 33.966
 }

gate4_label2id.json

@@ -0,0 +1,5 @@
+{
+  "O": 0,
+  "B-trigger": 1,
+  "I-trigger": 2
+}

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a3764aec317c8d33aa9d03c8e6c81eab6a79d179058f2c5d3b285e08b28f74e6
-size 433108512
+oid sha256:7f23fa81eeb6187cb809748a50fd157970d17bf2457c08bb2d53e62c79a3f94a
+size 433117924

test_results.json

@@ -1,22 +1,28 @@
 {
     "epoch": 10.0,
-    "eval_gate1_accuracy": 0.9204524721766101,
-    "eval_gate1_f1": 0.9175696677724912,
-    "eval_gate1_precision": 0.9198718411071631,
-    "eval_gate1_recall": 0.9156173695177233,
+    "eval_gate1_accuracy": 0.9250136836343733,
+    "eval_gate1_f1": 0.9224529503160914,
+    "eval_gate1_precision": 0.9237612665493629,
+    "eval_gate1_recall": 0.9212694184275014,
     "eval_gate1_support": 5481,
-    "eval_gate2_accuracy": 0.8573253056011677,
-    "eval_gate2_f1": 0.7662196765345953,
-    "eval_gate2_precision": 0.7540255243234163,
-    "eval_gate2_recall": 0.7899899686908007,
+    "eval_gate2_accuracy": 0.8680897646414888,
+    "eval_gate2_f1": 0.8167601906166684,
+    "eval_gate2_precision": 0.8143474395151585,
+    "eval_gate2_recall": 0.8232715925733067,
     "eval_gate2_support": 5481,
-    "eval_gate3_accuracy": 0.8761520737327189,
-    "eval_gate3_f1": 0.7204269422018509,
-    "eval_gate3_precision": 0.7065840948237819,
-    "eval_gate3_recall": 0.7411799394675562,
+    "eval_gate3_accuracy": 0.8945852534562212,
+    "eval_gate3_f1": 0.814049046653749,
+    "eval_gate3_precision": 0.8216561919229995,
+    "eval_gate3_recall": 0.8169058833981815,
     "eval_gate3_support": 6944,
-    "eval_loss": 0.7167699933052063,
-    "eval_runtime": 9.951,
-    "eval_samples_per_second": 1248.614,
-    "eval_steps_per_second": 39.091
+    "eval_gate4_span_FN": 200,
+    "eval_gate4_span_FP": 210,
+    "eval_gate4_span_TP": 811,
+    "eval_gate4_span_f1": 0.7982283414489575,
+    "eval_gate4_span_precision": 0.794319294801231,
+    "eval_gate4_span_recall": 0.8021760632957252,
+    "eval_loss": NaN,
+    "eval_runtime": 12.712,
+    "eval_samples_per_second": 1057.505,
+    "eval_steps_per_second": 33.118
 }

trainer_state.json

@@ -1,372 +1,444 @@
 {
-  "best_global_step": 15535,
-  "best_metric": 0.91874922234221,
-  "best_model_checkpoint": "output/indra_stmt_classifier/checkpoint-15535",
+  "best_global_step": 33610,
+  "best_metric": 0.922882664177459,
+  "best_model_checkpoint": "output/indra_stmt_classifier/checkpoint-33610",
   "epoch": 10.0,
   "eval_steps": 500,
-  "global_step": 31070,
+  "global_step": 33610,
   "is_hyper_param_search": false,
   "is_local_process_zero": true,
   "is_world_process_zero": true,
   "log_history": [
     {
       "epoch": 1.0,
-      "grad_norm": 2.75587797164917,
-      "learning_rate": 2.7000965561635017e-05,
-      "loss": 1.1186,
-      "step": 3107
+      "grad_norm": 88.24068450927734,
+      "learning_rate": 2.7000892591490628e-05,
+      "loss": 2.1663,
+      "step": 3361
     },
     {
       "epoch": 1.0,
-      "eval_gate1_accuracy": 0.8848540145985402,
-      "eval_gate1_f1": 0.8785447299140416,
-      "eval_gate1_precision": 0.879748756806488,
-      "eval_gate1_recall": 0.8774234499312313,
+      "eval_gate1_accuracy": 0.8695255474452555,
+      "eval_gate1_f1": 0.8578099178217389,
+      "eval_gate1_precision": 0.8773808096928044,
+      "eval_gate1_recall": 0.8475299794051141,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.7870437956204379,
-      "eval_gate2_f1": 0.5722188113150324,
-      "eval_gate2_precision": 0.5654852652404123,
-      "eval_gate2_recall": 0.5860432299031674,
+      "eval_gate2_accuracy": 0.7768248175182482,
+      "eval_gate2_f1": 0.5680534916799017,
+      "eval_gate2_precision": 0.5577177732555311,
+      "eval_gate2_recall": 0.5921072842711521,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8329493087557603,
-      "eval_gate3_f1": 0.639519347317039,
-      "eval_gate3_precision": 0.677673907531002,
-      "eval_gate3_recall": 0.6297629647896671,
+      "eval_gate3_accuracy": 0.8647753456221198,
+      "eval_gate3_f1": 0.7215116868347058,
+      "eval_gate3_precision": 0.7091765529489763,
+      "eval_gate3_recall": 0.7377508812382757,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.7573269009590149,
-      "eval_runtime": 9.8025,
-      "eval_samples_per_second": 1267.429,
-      "eval_steps_per_second": 39.684,
-      "step": 3107
+      "eval_gate4_span_FN": 237,
+      "eval_gate4_span_FP": 284,
+      "eval_gate4_span_TP": 770,
+      "eval_gate4_span_f1": 0.7472100871836075,
+      "eval_gate4_span_precision": 0.7305502846230498,
+      "eval_gate4_span_recall": 0.7646474677183253,
+      "eval_loss": NaN,
+      "eval_runtime": 12.4516,
+      "eval_samples_per_second": 1079.542,
+      "eval_steps_per_second": 33.811,
+      "step": 3361
     },
     {
       "epoch": 2.0,
-      "grad_norm": 9.97591781616211,
-      "learning_rate": 2.4000965561635018e-05,
-      "loss": 0.6341,
-      "step": 6214
+      "grad_norm": 22.65692901611328,
+      "learning_rate": 2.400089259149063e-05,
+      "loss": 1.1836,
+      "step": 6722
     },
     {
       "epoch": 2.0,
-      "eval_gate1_accuracy": 0.8994525547445256,
-      "eval_gate1_f1": 0.8920945233221212,
-      "eval_gate1_precision": 0.9025475703450652,
-      "eval_gate1_recall": 0.8850382209313125,
+      "eval_gate1_accuracy": 0.8985401459854014,
+      "eval_gate1_f1": 0.891649060279065,
+      "eval_gate1_precision": 0.8991425334535017,
+      "eval_gate1_recall": 0.8861618139187699,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8173357664233577,
-      "eval_gate2_f1": 0.6503078056225045,
-      "eval_gate2_precision": 0.6679383283455959,
-      "eval_gate2_recall": 0.6597425206224358,
+      "eval_gate2_accuracy": 0.8175182481751825,
+      "eval_gate2_f1": 0.6618300467616971,
+      "eval_gate2_precision": 0.713493817631298,
+      "eval_gate2_recall": 0.6674431137299971,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8425979262672811,
-      "eval_gate3_f1": 0.7516576653132845,
-      "eval_gate3_precision": 0.7289310992313269,
-      "eval_gate3_recall": 0.7877551283882368,
+      "eval_gate3_accuracy": 0.8798963133640553,
+      "eval_gate3_f1": 0.7753981617797622,
+      "eval_gate3_precision": 0.7757031362148418,
+      "eval_gate3_recall": 0.7910325664330483,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.681019127368927,
-      "eval_runtime": 9.815,
-      "eval_samples_per_second": 1265.816,
-      "eval_steps_per_second": 39.633,
-      "step": 6214
+      "eval_gate4_span_FN": 276,
+      "eval_gate4_span_FP": 195,
+      "eval_gate4_span_TP": 731,
+      "eval_gate4_span_f1": 0.7563372945353566,
+      "eval_gate4_span_precision": 0.7894168466437428,
+      "eval_gate4_span_recall": 0.7259185700027218,
+      "eval_loss": NaN,
+      "eval_runtime": 12.374,
+      "eval_samples_per_second": 1086.309,
+      "eval_steps_per_second": 34.023,
+      "step": 6722
     },
     {
       "epoch": 3.0,
-      "grad_norm": 6.9430036544799805,
-      "learning_rate": 2.100096556163502e-05,
-      "loss": 0.4397,
-      "step": 9321
+      "grad_norm": 48.53071975708008,
+      "learning_rate": 2.1000892591490628e-05,
+      "loss": 0.8284,
+      "step": 10083
     },
     {
       "epoch": 3.0,
-      "eval_gate1_accuracy": 0.9094890510948905,
-      "eval_gate1_f1": 0.9044200252650033,
-      "eval_gate1_precision": 0.9062240634591001,
-      "eval_gate1_recall": 0.90277723499615,
+      "eval_gate1_accuracy": 0.9155109489051095,
+      "eval_gate1_f1": 0.9106916043370983,
+      "eval_gate1_precision": 0.9129870935118751,
+      "eval_gate1_recall": 0.908642288137623,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.839051094890511,
-      "eval_gate2_f1": 0.7263650484208171,
-      "eval_gate2_precision": 0.7759514146848825,
-      "eval_gate2_recall": 0.7203927286117487,
+      "eval_gate2_accuracy": 0.8507299270072993,
+      "eval_gate2_f1": 0.778394417869388,
+      "eval_gate2_precision": 0.7929090995262216,
+      "eval_gate2_recall": 0.7749487518215912,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8652073732718893,
-      "eval_gate3_f1": 0.7603754676303877,
-      "eval_gate3_precision": 0.733971679937432,
-      "eval_gate3_recall": 0.8035307301942258,
+      "eval_gate3_accuracy": 0.8847926267281107,
+      "eval_gate3_f1": 0.7812164185694548,
+      "eval_gate3_precision": 0.7677202242109172,
+      "eval_gate3_recall": 0.8143775059784232,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.6797090172767639,
-      "eval_runtime": 9.843,
-      "eval_samples_per_second": 1262.211,
-      "eval_steps_per_second": 39.52,
-      "step": 9321
+      "eval_gate4_span_FN": 237,
+      "eval_gate4_span_FP": 213,
+      "eval_gate4_span_TP": 770,
+      "eval_gate4_span_f1": 0.7738693417266181,
+      "eval_gate4_span_precision": 0.7833163784253986,
+      "eval_gate4_span_recall": 0.7646474677183253,
+      "eval_loss": NaN,
+      "eval_runtime": 12.5039,
+      "eval_samples_per_second": 1075.022,
+      "eval_steps_per_second": 33.669,
+      "step": 10083
     },
     {
       "epoch": 4.0,
-      "grad_norm": 2.871424913406372,
-      "learning_rate": 1.8000965561635018e-05,
-      "loss": 0.3263,
-      "step": 12428
+      "grad_norm": 28.591859817504883,
+      "learning_rate": 1.800089259149063e-05,
+      "loss": 0.6053,
+      "step": 13444
     },
     {
       "epoch": 4.0,
-      "eval_gate1_accuracy": 0.9138686131386862,
-      "eval_gate1_f1": 0.9095152677593289,
-      "eval_gate1_precision": 0.9091069849690538,
-      "eval_gate1_recall": 0.909933548120778,
+      "eval_gate1_accuracy": 0.9187956204379562,
+      "eval_gate1_f1": 0.9134909277042625,
+      "eval_gate1_precision": 0.9199399576679961,
+      "eval_gate1_recall": 0.9085254675240795,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8523722627737226,
-      "eval_gate2_f1": 0.7611458139063093,
-      "eval_gate2_precision": 0.8024369787744103,
-      "eval_gate2_recall": 0.7429072801122445,
+      "eval_gate2_accuracy": 0.8541970802919708,
+      "eval_gate2_f1": 0.7744618041449975,
+      "eval_gate2_precision": 0.7571752325760932,
+      "eval_gate2_recall": 0.8029872948457836,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8620391705069125,
-      "eval_gate3_f1": 0.7744302908340429,
-      "eval_gate3_precision": 0.7519119769092221,
-      "eval_gate3_recall": 0.811903740069826,
+      "eval_gate3_accuracy": 0.8696716589861752,
+      "eval_gate3_f1": 0.7354622549959394,
+      "eval_gate3_precision": 0.6870145918379549,
+      "eval_gate3_recall": 0.8067201505001067,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.6769687533378601,
-      "eval_runtime": 9.7569,
-      "eval_samples_per_second": 1273.351,
-      "eval_steps_per_second": 39.869,
-      "step": 12428
+      "eval_gate4_span_FN": 221,
+      "eval_gate4_span_FP": 226,
+      "eval_gate4_span_TP": 786,
+      "eval_gate4_span_f1": 0.7786032639373403,
+      "eval_gate4_span_precision": 0.7766798418895585,
+      "eval_gate4_span_recall": 0.7805362462683164,
+      "eval_loss": NaN,
+      "eval_runtime": 12.4674,
+      "eval_samples_per_second": 1078.168,
+      "eval_steps_per_second": 33.768,
+      "step": 13444
     },
     {
       "epoch": 5.0,
-      "grad_norm": 5.60059928894043,
-      "learning_rate": 1.5000965561635018e-05,
-      "loss": 0.2441,
-      "step": 15535
+      "grad_norm": 33.3723030090332,
+      "learning_rate": 1.5000892591490627e-05,
+      "loss": 0.4558,
+      "step": 16805
     },
     {
       "epoch": 5.0,
-      "eval_gate1_accuracy": 0.9231751824817518,
-      "eval_gate1_f1": 0.91874922234221,
-      "eval_gate1_precision": 0.9213536363863121,
-      "eval_gate1_recall": 0.916446997689221,
+      "eval_gate1_accuracy": 0.9198905109489051,
+      "eval_gate1_f1": 0.9149163874867263,
+      "eval_gate1_precision": 0.9196832242670131,
+      "eval_gate1_recall": 0.9110372507879438,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8593065693430657,
-      "eval_gate2_f1": 0.7763067392280264,
-      "eval_gate2_precision": 0.7921477517391428,
-      "eval_gate2_recall": 0.775442797363134,
+      "eval_gate2_accuracy": 0.8604014598540146,
+      "eval_gate2_f1": 0.793154234054575,
+      "eval_gate2_precision": 0.7914032181614735,
+      "eval_gate2_recall": 0.798016426758543,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8810483870967742,
-      "eval_gate3_f1": 0.778218086166726,
-      "eval_gate3_precision": 0.75635112362015,
-      "eval_gate3_recall": 0.8147429373450497,
+      "eval_gate3_accuracy": 0.8922811059907834,
+      "eval_gate3_f1": 0.7877507428253906,
+      "eval_gate3_precision": 0.7734088072287458,
+      "eval_gate3_recall": 0.8143664133743507,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.7152989506721497,
-      "eval_runtime": 9.6947,
-      "eval_samples_per_second": 1281.524,
-      "eval_steps_per_second": 40.125,
-      "step": 15535
+      "eval_gate4_span_FN": 209,
+      "eval_gate4_span_FP": 199,
+      "eval_gate4_span_TP": 798,
+      "eval_gate4_span_f1": 0.796407180620919,
+      "eval_gate4_span_precision": 0.8004012036028044,
+      "eval_gate4_span_recall": 0.7924528301808098,
+      "eval_loss": NaN,
+      "eval_runtime": 12.3943,
+      "eval_samples_per_second": 1084.529,
+      "eval_steps_per_second": 33.967,
+      "step": 16805
     },
     {
       "epoch": 6.0,
-      "grad_norm": 16.034574508666992,
-      "learning_rate": 1.2000965561635018e-05,
-      "loss": 0.1855,
-      "step": 18642
+      "grad_norm": 8.42479133605957,
+      "learning_rate": 1.2000892591490627e-05,
+      "loss": 0.3525,
+      "step": 20166
     },
     {
       "epoch": 6.0,
-      "eval_gate1_accuracy": 0.920985401459854,
-      "eval_gate1_f1": 0.9165971625446729,
-      "eval_gate1_precision": 0.9182460374588926,
-      "eval_gate1_recall": 0.9150794681328354,
+      "eval_gate1_accuracy": 0.9224452554744526,
+      "eval_gate1_f1": 0.9184357453857448,
+      "eval_gate1_precision": 0.9184713900488881,
+      "eval_gate1_recall": 0.9184001711127836,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.862956204379562,
-      "eval_gate2_f1": 0.7924841981022327,
-      "eval_gate2_precision": 0.8068279196629584,
-      "eval_gate2_recall": 0.7874456055649913,
+      "eval_gate2_accuracy": 0.8642335766423358,
+      "eval_gate2_f1": 0.8044365885423028,
+      "eval_gate2_precision": 0.8273316033708245,
+      "eval_gate2_recall": 0.7886231573533272,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8882488479262672,
-      "eval_gate3_f1": 0.7926910665874005,
-      "eval_gate3_precision": 0.7755825529327147,
-      "eval_gate3_recall": 0.8216291424595251,
+      "eval_gate3_accuracy": 0.8919930875576036,
+      "eval_gate3_f1": 0.7862482126489905,
+      "eval_gate3_precision": 0.7715028053430636,
+      "eval_gate3_recall": 0.813526414966601,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.8007788062095642,
-      "eval_runtime": 9.803,
-      "eval_samples_per_second": 1267.368,
-      "eval_steps_per_second": 39.682,
-      "step": 18642
+      "eval_gate4_span_FN": 214,
+      "eval_gate4_span_FP": 247,
+      "eval_gate4_span_TP": 793,
+      "eval_gate4_span_f1": 0.7747923740850827,
+      "eval_gate4_span_precision": 0.7624999999926684,
+      "eval_gate4_span_recall": 0.7874875868839376,
+      "eval_loss": NaN,
+      "eval_runtime": 12.3837,
+      "eval_samples_per_second": 1085.458,
+      "eval_steps_per_second": 33.996,
+      "step": 20166
     },
     {
       "epoch": 7.0,
-      "grad_norm": 11.89713191986084,
-      "learning_rate": 9.000965561635018e-06,
-      "loss": 0.1454,
-      "step": 21749
+      "grad_norm": 40.48398208618164,
+      "learning_rate": 9.000892591490627e-06,
+      "loss": 0.273,
+      "step": 23527
     },
     {
       "epoch": 7.0,
-      "eval_gate1_accuracy": 0.9193430656934306,
-      "eval_gate1_f1": 0.9147193543716858,
-      "eval_gate1_precision": 0.9171674547406892,
-      "eval_gate1_recall": 0.9125446429134221,
+      "eval_gate1_accuracy": 0.9264598540145985,
+      "eval_gate1_f1": 0.9225784404073203,
+      "eval_gate1_precision": 0.9230570530098832,
+      "eval_gate1_recall": 0.9221118169373361,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8640510948905109,
-      "eval_gate2_f1": 0.7946506784893311,
-      "eval_gate2_precision": 0.8100667075746566,
-      "eval_gate2_recall": 0.7852244900603023,
+      "eval_gate2_accuracy": 0.8728102189781022,
+      "eval_gate2_f1": 0.8041274676974263,
+      "eval_gate2_precision": 0.8118073730537297,
+      "eval_gate2_recall": 0.8000902844697497,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8895449308755761,
-      "eval_gate3_f1": 0.7920894399575168,
-      "eval_gate3_precision": 0.7849169985769698,
-      "eval_gate3_recall": 0.8118891174634335,
+      "eval_gate3_accuracy": 0.8930011520737328,
+      "eval_gate3_f1": 0.7925867281754787,
+      "eval_gate3_precision": 0.7856067431455332,
+      "eval_gate3_recall": 0.8106698879350709,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.8615767955780029,
-      "eval_runtime": 9.8147,
-      "eval_samples_per_second": 1265.852,
-      "eval_steps_per_second": 39.634,
-      "step": 21749
+      "eval_gate4_span_FN": 199,
+      "eval_gate4_span_FP": 235,
+      "eval_gate4_span_TP": 808,
+      "eval_gate4_span_f1": 0.7882926779206807,
+      "eval_gate4_span_precision": 0.7746883988420452,
+      "eval_gate4_span_recall": 0.8023833167745543,
+      "eval_loss": NaN,
+      "eval_runtime": 12.4014,
+      "eval_samples_per_second": 1083.91,
+      "eval_steps_per_second": 33.948,
+      "step": 23527
     },
     {
       "epoch": 8.0,
-      "grad_norm": 0.0030767242424190044,
-      "learning_rate": 6.000965561635018e-06,
-      "loss": 0.1081,
-      "step": 24856
+      "grad_norm": 0.31248366832733154,
+      "learning_rate": 6.000892591490628e-06,
+      "loss": 0.21,
+      "step": 26888
     },
     {
       "epoch": 8.0,
-      "eval_gate1_accuracy": 0.9198905109489051,
-      "eval_gate1_f1": 0.9149957554765835,
-      "eval_gate1_precision": 0.9192622220801286,
-      "eval_gate1_recall": 0.9114623713660069,
+      "eval_gate1_accuracy": 0.9257299270072993,
+      "eval_gate1_f1": 0.9216466278937456,
+      "eval_gate1_precision": 0.9230669021709708,
+      "eval_gate1_recall": 0.9203239292739431,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8638686131386861,
-      "eval_gate2_f1": 0.7968208804038338,
-      "eval_gate2_precision": 0.7930637027689245,
-      "eval_gate2_recall": 0.8065230155153877,
+      "eval_gate2_accuracy": 0.8706204379562044,
+      "eval_gate2_f1": 0.8069785114481128,
+      "eval_gate2_precision": 0.8101824567894452,
+      "eval_gate2_recall": 0.8050611222862718,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8922811059907834,
-      "eval_gate3_f1": 0.8026929577988032,
-      "eval_gate3_precision": 0.7897651226837507,
-      "eval_gate3_recall": 0.8257808045726754,
+      "eval_gate3_accuracy": 0.8967453917050692,
+      "eval_gate3_f1": 0.7998434306583326,
+      "eval_gate3_precision": 0.7931305927714913,
+      "eval_gate3_recall": 0.8165349639656843,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.9336212873458862,
-      "eval_runtime": 9.7984,
-      "eval_samples_per_second": 1267.963,
-      "eval_steps_per_second": 39.7,
-      "step": 24856
+      "eval_gate4_span_FN": 212,
+      "eval_gate4_span_FP": 202,
+      "eval_gate4_span_TP": 795,
+      "eval_gate4_span_f1": 0.793413168644901,
+      "eval_gate4_span_precision": 0.7973921765215909,
+      "eval_gate4_span_recall": 0.7894736842026865,
+      "eval_loss": NaN,
+      "eval_runtime": 12.456,
+      "eval_samples_per_second": 1079.158,
+      "eval_steps_per_second": 33.799,
+      "step": 26888
     },
     {
       "epoch": 9.0,
-      "grad_norm": 116.7902603149414,
-      "learning_rate": 3.0009655616350177e-06,
-      "loss": 0.0842,
-      "step": 27963
+      "grad_norm": 99.82740020751953,
+      "learning_rate": 3.0008925914906276e-06,
+      "loss": 0.1662,
+      "step": 30249
     },
     {
       "epoch": 9.0,
-      "eval_gate1_accuracy": 0.9215328467153284,
-      "eval_gate1_f1": 0.9169295960347313,
-      "eval_gate1_precision": 0.9200202922521744,
-      "eval_gate1_recall": 0.9142522689322582,
+      "eval_gate1_accuracy": 0.9253649635036496,
+      "eval_gate1_f1": 0.921206016357208,
+      "eval_gate1_precision": 0.9229565396154125,
+      "eval_gate1_recall": 0.9196000336734718,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8656934306569343,
-      "eval_gate2_f1": 0.7916047345282854,
-      "eval_gate2_precision": 0.7975761529206665,
-      "eval_gate2_recall": 0.7905660788693317,
+      "eval_gate2_accuracy": 0.8735401459854014,
+      "eval_gate2_f1": 0.8113461645918237,
+      "eval_gate2_precision": 0.8191561758392163,
+      "eval_gate2_recall": 0.8066820228163906,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8908410138248848,
-      "eval_gate3_f1": 0.7962561702350882,
-      "eval_gate3_precision": 0.7845233050024135,
-      "eval_gate3_recall": 0.8178146646821465,
+      "eval_gate3_accuracy": 0.8963133640552995,
+      "eval_gate3_f1": 0.7979599095284565,
+      "eval_gate3_precision": 0.7889647625019129,
+      "eval_gate3_recall": 0.8173306025690692,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.9843125939369202,
-      "eval_runtime": 9.6312,
-      "eval_samples_per_second": 1289.977,
-      "eval_steps_per_second": 40.39,
-      "step": 27963
+      "eval_gate4_span_FN": 194,
+      "eval_gate4_span_FP": 204,
+      "eval_gate4_span_TP": 813,
+      "eval_gate4_span_f1": 0.8033596787866502,
+      "eval_gate4_span_precision": 0.7994100294906646,
+      "eval_gate4_span_recall": 0.8073485600714265,
+      "eval_loss": NaN,
+      "eval_runtime": 12.4958,
+      "eval_samples_per_second": 1075.724,
+      "eval_steps_per_second": 33.691,
+      "step": 30249
     },
     {
       "epoch": 10.0,
-      "grad_norm": 0.09423983097076416,
-      "learning_rate": 9.655616350177021e-10,
-      "loss": 0.068,
-      "step": 31070
+      "grad_norm": 15.471023559570312,
+      "learning_rate": 8.925914906277894e-10,
+      "loss": 0.1297,
+      "step": 33610
     },
     {
       "epoch": 10.0,
-      "eval_gate1_accuracy": 0.920985401459854,
-      "eval_gate1_f1": 0.916357639651535,
-      "eval_gate1_precision": 0.9193983763664938,
-      "eval_gate1_recall": 0.9137190822830333,
+      "eval_gate1_accuracy": 0.9268248175182482,
+      "eval_gate1_f1": 0.922882664177459,
+      "eval_gate1_precision": 0.9238243397462302,
+      "eval_gate1_recall": 0.9219854713816811,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8633211678832117,
-      "eval_gate2_f1": 0.7921038946740504,
-      "eval_gate2_precision": 0.802926749773104,
-      "eval_gate2_recall": 0.7855851424596473,
+      "eval_gate2_accuracy": 0.8728102189781022,
+      "eval_gate2_f1": 0.7972864171672726,
+      "eval_gate2_precision": 0.8049759563676634,
+      "eval_gate2_recall": 0.7924699515870988,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8915610599078341,
-      "eval_gate3_f1": 0.8022192601674913,
-      "eval_gate3_precision": 0.7867252072690095,
-      "eval_gate3_recall": 0.8274418766924675,
+      "eval_gate3_accuracy": 0.8955933179723502,
+      "eval_gate3_f1": 0.7963151574821563,
+      "eval_gate3_precision": 0.7875457145835848,
+      "eval_gate3_recall": 0.8152968118858834,
       "eval_gate3_support": 6944,
-      "eval_loss": 1.0077552795410156,
-      "eval_runtime": 9.7597,
-      "eval_samples_per_second": 1272.994,
-      "eval_steps_per_second": 39.858,
-      "step": 31070
+      "eval_gate4_span_FN": 196,
+      "eval_gate4_span_FP": 216,
+      "eval_gate4_span_TP": 811,
+      "eval_gate4_span_f1": 0.7974434561529178,
+      "eval_gate4_span_precision": 0.7896786757469361,
+      "eval_gate4_span_recall": 0.8053624627526776,
+      "eval_loss": NaN,
+      "eval_runtime": 12.516,
+      "eval_samples_per_second": 1073.986,
+      "eval_steps_per_second": 33.637,
+      "step": 33610
     },
     {
       "epoch": 10.0,
-      "step": 31070,
-      "total_flos": 6.336727068480365e+16,
-      "train_loss": 0.3354094574986296,
-      "train_runtime": 2265.1995,
-      "train_samples_per_second": 438.787,
-      "train_steps_per_second": 13.716
+      "step": 33610,
+      "total_flos": 6.790211473893735e+16,
+      "train_loss": 0.637072107416645,
+      "train_runtime": 2571.8808,
+      "train_samples_per_second": 418.118,
+      "train_steps_per_second": 13.068
     },
     {
       "epoch": 10.0,
-      "eval_gate1_accuracy": 0.9231751824817518,
-      "eval_gate1_f1": 0.91874922234221,
-      "eval_gate1_precision": 0.9213536363863121,
-      "eval_gate1_recall": 0.916446997689221,
+      "eval_gate1_accuracy": 0.9268248175182482,
+      "eval_gate1_f1": 0.922882664177459,
+      "eval_gate1_precision": 0.9238243397462302,
+      "eval_gate1_recall": 0.9219854713816811,
       "eval_gate1_support": 5480,
-      "eval_gate2_accuracy": 0.8593065693430657,
-      "eval_gate2_f1": 0.7763067392280264,
-      "eval_gate2_precision": 0.7921477517391428,
-      "eval_gate2_recall": 0.775442797363134,
+      "eval_gate2_accuracy": 0.8728102189781022,
+      "eval_gate2_f1": 0.7972864171672726,
+      "eval_gate2_precision": 0.8049759563676634,
+      "eval_gate2_recall": 0.7924699515870988,
       "eval_gate2_support": 5480,
-      "eval_gate3_accuracy": 0.8810483870967742,
-      "eval_gate3_f1": 0.778218086166726,
-      "eval_gate3_precision": 0.75635112362015,
-      "eval_gate3_recall": 0.8147429373450497,
+      "eval_gate3_accuracy": 0.8955933179723502,
+      "eval_gate3_f1": 0.7963151574821563,
+      "eval_gate3_precision": 0.7875457145835848,
+      "eval_gate3_recall": 0.8152968118858834,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.7152989506721497,
-      "eval_runtime": 9.7809,
-      "eval_samples_per_second": 1270.237,
-      "eval_steps_per_second": 39.772,
-      "step": 31070
+      "eval_gate4_span_FN": 196,
+      "eval_gate4_span_FP": 216,
+      "eval_gate4_span_TP": 811,
+      "eval_gate4_span_f1": 0.7974434561529178,
+      "eval_gate4_span_precision": 0.7896786757469361,
+      "eval_gate4_span_recall": 0.8053624627526776,
+      "eval_loss": NaN,
+      "eval_runtime": 12.3946,
+      "eval_samples_per_second": 1084.506,
+      "eval_steps_per_second": 33.966,
+      "step": 33610
     },
     {
       "epoch": 10.0,
-      "eval_gate1_accuracy": 0.9204524721766101,
-      "eval_gate1_f1": 0.9175696677724912,
-      "eval_gate1_precision": 0.9198718411071631,
-      "eval_gate1_recall": 0.9156173695177233,
+      "eval_gate1_accuracy": 0.9250136836343733,
+      "eval_gate1_f1": 0.9224529503160914,
+      "eval_gate1_precision": 0.9237612665493629,
+      "eval_gate1_recall": 0.9212694184275014,
       "eval_gate1_support": 5481,
-      "eval_gate2_accuracy": 0.8573253056011677,
-      "eval_gate2_f1": 0.7662196765345953,
-      "eval_gate2_precision": 0.7540255243234163,
-      "eval_gate2_recall": 0.7899899686908007,
+      "eval_gate2_accuracy": 0.8680897646414888,
+      "eval_gate2_f1": 0.8167601906166684,
+      "eval_gate2_precision": 0.8143474395151585,
+      "eval_gate2_recall": 0.8232715925733067,
       "eval_gate2_support": 5481,
-      "eval_gate3_accuracy": 0.8761520737327189,
-      "eval_gate3_f1": 0.7204269422018509,
-      "eval_gate3_precision": 0.7065840948237819,
-      "eval_gate3_recall": 0.7411799394675562,
+      "eval_gate3_accuracy": 0.8945852534562212,
+      "eval_gate3_f1": 0.814049046653749,
+      "eval_gate3_precision": 0.8216561919229995,
+      "eval_gate3_recall": 0.8169058833981815,
       "eval_gate3_support": 6944,
-      "eval_loss": 0.7167699933052063,
-      "eval_runtime": 9.951,
-      "eval_samples_per_second": 1248.614,
-      "eval_steps_per_second": 39.091,
-      "step": 31070
+      "eval_gate4_span_FN": 200,
+      "eval_gate4_span_FP": 210,
+      "eval_gate4_span_TP": 811,
+      "eval_gate4_span_f1": 0.7982283414489575,
+      "eval_gate4_span_precision": 0.794319294801231,
+      "eval_gate4_span_recall": 0.8021760632957252,
+      "eval_loss": NaN,
+      "eval_runtime": 12.712,
+      "eval_samples_per_second": 1057.505,
+      "eval_steps_per_second": 33.118,
+      "step": 33610
     }
   ],
   "logging_steps": 500,
-  "max_steps": 31070,
+  "max_steps": 33610,
   "num_input_tokens_seen": 0,
   "num_train_epochs": 10,
   "save_steps": 500,
@@ -382,7 +454,7 @@
       "attributes": {}
     }
   },
-  "total_flos": 6.336727068480365e+16,
+  "total_flos": 6.790211473893735e+16,
   "train_batch_size": 32,
   "trial_name": null,
   "trial_params": null

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2f7fd5367a51397aa9e61e2f6c06b3f993347adf643ea400b270a4147e24f944
+oid sha256:db7d845da31495446865a3d678a3acc159a3d3032d0d5a58fc1e1da4a4289222
 size 5841