gloamxun/modeling_zhaocui
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

更新 math Q1 Q2 实验与特征

Browse Source
This commit is contained in:
gloamxun
2026-09-25 08:58:10 +08:00
parent ca5be95757
commit 83ec3d1a83
168 changed files with 150303 additions and 2874 deletions
Download Patch File Download Diff File Expand all files Collapse all files
math/Q2/results/ablation_validation.csv
+13
View File
@@ -0,0 +1,13 @@
model,n,accuracy,macro_f1,negative_support,neutral_support,positive_support,negative_recall,middle_recall,positive_recall,regression_mae,regression_rmse,pearson,brier,classification_nll,ece_15,interval_90_coverage,interval_90_mean_width,selection_nll,temperature,description,predictive_variance_mean_uncalibrated,within_trajectory_variance_mean,between_trajectory_variance_mean,predictive_mean_mean_calibrated,predictive_variance_mean_calibrated,validation_selection_loss,rho_imp,lambda_u,lambda_gap,lambda_span,lambda_group,group_temperature
C0,728,0.592032967032967,0.5559346126510305,206,184,338,0.5825242718446602,0.34782608695652173,0.7307692307692307,0.7334139347076416,0.9371736594281902,0.5192425847053528,0.5145634913739343,0.8668775768593128,0.03562808104126608,0.9093406593406593,3.1437528133392334,0.9035571651175792,2.032287887301264,observed mean/std + masks + maximum gap; logistic/ridge,,,,,,,,,,,,
C1,728,0.6071428571428571,0.55124288334091,206,184,338,0.6796116504854369,0.23369565217391305,0.7662721893491125,0.6779899001121521,0.9222378420815579,0.5779994130134583,0.4974337085664247,0.8442830443382263,0.05466654393207895,0.8956043956043956,2.463564872741699,2.7879433631896973,1.1214790595552069,"masked BiGRU; no posterior imputation, explicit reliability or source gate",0.5905959010124207,0.5905959010124207,0.0,0.18120460212230682,0.6191508769989014,2.7879433631896973,0.5,0.05,0.05,0.05,nan,nan
C2,728,0.5975274725274725,0.47637082507472805,206,184,338,0.6116504854368932,0.04891304347826087,0.8875739644970414,0.7159271240234375,0.9473113058814692,0.579660177230835,0.5006998592970549,0.8497811555862427,0.07025201639140045,0.8983516483516484,2.5833804607391357,2.7831006050109863,1.2159886200859098,exact Gaussian posterior mean; no joint trajectory integral,0.6255680322647095,0.6255680322647095,0.0,0.26571935415267944,0.6729797124862671,2.7831006050109863,0.5,0.05,0.05,0.05,nan,nan
C3,728,0.6085164835164835,0.4974009120786063,206,184,338,0.6553398058252428,0.07608695652173914,0.8698224852071006,0.7170785665512085,0.9519565799792883,0.5966356992721558,0.4963717223322455,0.8414766192436218,0.03515750726500711,0.9065934065934066,2.5035855770111084,2.771331548690796,1.2137760834757865,joint trajectory integral plus final reliability/content fusion gate,0.5990039110183716,0.5990003943443298,3.476098072496825e-06,0.21980196237564087,0.6516103148460388,2.77126407623291,0.3,0.05,0.05,0.05,nan,nan
C4,728,0.6085164835164835,0.555058605423057,206,184,338,0.6504854368932039,0.25,0.7781065088757396,0.6918120384216309,0.9515582480750588,0.5684906244277954,0.5075791281376366,0.8570581078529358,0.05379025265574453,0.8983516483516484,2.538222312927246,2.8107874393463135,1.0471160328013351,C3 plus bounded cross-time source attention and null source,0.6382710933685303,0.6382405161857605,3.0667371902382e-05,0.2080966681241989,0.6502240300178528,2.81081223487854,0.7,0.05,0.05,0.05,nan,nan
C5,728,0.6153846153846154,0.5445555339287688,206,184,338,0.6990291262135923,0.19021739130434784,0.7958579881656804,0.6623618006706238,0.8933437077491609,0.6174448132514954,0.4918066281209682,0.8373782634735107,0.06113330221601894,0.8928571428571429,2.381938934326172,2.7620162963867188,1.212728800581531,C4 plus reliability-modulated BiGRU update,0.5485545992851257,0.5485460758209229,8.503861863573547e-06,0.13911069929599762,0.5960606336593628,2.7618556022644043,0.3,0.05,0.05,0.05,nan,nan
C6,728,0.6263736263736264,0.5801114274083624,206,184,338,0.6941747572815534,0.29891304347826086,0.7633136094674556,0.6457235217094421,0.8814069986838935,0.6157047748565674,0.4856194350316264,0.8258335590362549,0.031823099805758544,0.885989010989011,2.30387282371521,2.774562358856201,1.1780751742536886,C5 plus optional rank-4 CP residual,0.49419134855270386,0.4941462278366089,4.511491351877339e-05,0.13551649451255798,0.5317712426185608,2.774617910385132,0.3,0.05,0.05,0.05,nan,nan
C6_no_distance,728,0.6043956043956044,0.5144916075030433,206,184,338,0.6553398058252428,0.13043478260869565,0.8313609467455622,0.6794722676277161,0.9098390018130259,0.6045995950698853,0.5041281436421426,0.8588091731071472,0.04508186811274226,0.885989010989011,2.2977938652038574,2.808736562728882,1.36875500088095,C6 with uncertainty retained but both distance/span reliability penalties fixed to zero,0.4714658260345459,0.47136420011520386,0.00010164274135604501,0.14785172045230865,0.5429188013076782,2.8084704875946045,0.5,0.0,0.0,0.0,nan,nan
C6_no_reconstruction,728,0.6002747252747253,0.5178357887707189,206,184,338,0.6650485436893204,0.14130434782608695,0.8106508875739645,0.6957471370697021,0.9327064671148756,0.5887437462806702,0.5024915749758389,0.8508719205856323,0.04982036520000343,0.8928571428571429,2.390105724334717,2.794447422027588,1.180521560150874,C6 trained without the auxiliary hidden-feature reconstruction loss,0.5634137392044067,0.5633965134620667,1.724438880046364e-05,0.1889886111021042,0.6063413023948669,2.794414520263672,0.3,0.05,0.05,0.05,nan,nan
C6_pointmask,728,0.603021978021978,0.5513375927177999,206,184,338,0.587378640776699,0.2717391304347826,0.7928994082840237,0.6945403218269348,0.9553973516704227,0.5646858811378479,0.49644123517098043,0.8433728814125061,0.04921124482547844,0.8997252747252747,2.4906656742095947,2.7908360958099365,1.3967463290480098,C6 trained with independent point masking instead of contiguous spans,0.5435255169868469,0.543519139289856,6.390413091139635e-06,0.234638050198555,0.6392570734024048,2.790741205215454,0.3,0.05,0.05,0.05,nan,nan
C7_distill,728,0.5906593406593407,0.5243846543289639,206,184,338,0.6601941747572816,0.18478260869565216,0.7692307692307693,0.6742856502532959,0.8981318409977092,0.6013374924659729,0.5035731740308492,0.8561364412307739,0.045249516246738015,0.8667582417582418,2.3018338680267334,2.8153398036956787,1.1711189639334147,C6 plus entropy/retention-weighted teacher distillation only,0.5134512186050415,0.513433575630188,1.761062230798416e-05,0.16756068170070648,0.5485631823539734,2.8153774738311768,0.3,0.05,0.05,0.05,nan,nan
C7_group,728,0.6085164835164835,0.5116842947005743,206,184,338,0.6747572815533981,0.11413043478260869,0.8372781065088757,0.6860273480415344,0.9252878475802377,0.5911917090415955,0.5028568696567639,0.8551272749900818,0.0504213785650311,0.8928571428571429,2.445509672164917,2.7835769653320312,1.044649831831679,C6 plus smooth worst-group risk only,0.6084297895431519,0.6084184646606445,1.1323560102027841e-05,0.12680216133594513,0.6198152303695679,2.783574342727661,0.3,0.05,0.05,0.05,0.1,0.05
1 model n accuracy macro_f1 negative_support neutral_support positive_support negative_recall middle_recall positive_recall regression_mae regression_rmse pearson brier classification_nll ece_15 interval_90_coverage interval_90_mean_width selection_nll temperature description predictive_variance_mean_uncalibrated within_trajectory_variance_mean between_trajectory_variance_mean predictive_mean_mean_calibrated predictive_variance_mean_calibrated validation_selection_loss rho_imp lambda_u lambda_gap lambda_span lambda_group group_temperature
2 C0 728 0.592032967032967 0.5559346126510305 206 184 338 0.5825242718446602 0.34782608695652173 0.7307692307692307 0.7334139347076416 0.9371736594281902 0.5192425847053528 0.5145634913739343 0.8668775768593128 0.03562808104126608 0.9093406593406593 3.1437528133392334 0.9035571651175792 2.032287887301264 observed mean/std + masks + maximum gap; logistic/ridge
3 C1 728 0.6071428571428571 0.55124288334091 206 184 338 0.6796116504854369 0.23369565217391305 0.7662721893491125 0.6779899001121521 0.9222378420815579 0.5779994130134583 0.4974337085664247 0.8442830443382263 0.05466654393207895 0.8956043956043956 2.463564872741699 2.7879433631896973 1.1214790595552069 masked BiGRU; no posterior imputation, explicit reliability or source gate 0.5905959010124207 0.5905959010124207 0.0 0.18120460212230682 0.6191508769989014 2.7879433631896973 0.5 0.05 0.05 0.05 nan nan
4 C2 728 0.5975274725274725 0.47637082507472805 206 184 338 0.6116504854368932 0.04891304347826087 0.8875739644970414 0.7159271240234375 0.9473113058814692 0.579660177230835 0.5006998592970549 0.8497811555862427 0.07025201639140045 0.8983516483516484 2.5833804607391357 2.7831006050109863 1.2159886200859098 exact Gaussian posterior mean; no joint trajectory integral 0.6255680322647095 0.6255680322647095 0.0 0.26571935415267944 0.6729797124862671 2.7831006050109863 0.5 0.05 0.05 0.05 nan nan
5 C3 728 0.6085164835164835 0.4974009120786063 206 184 338 0.6553398058252428 0.07608695652173914 0.8698224852071006 0.7170785665512085 0.9519565799792883 0.5966356992721558 0.4963717223322455 0.8414766192436218 0.03515750726500711 0.9065934065934066 2.5035855770111084 2.771331548690796 1.2137760834757865 joint trajectory integral plus final reliability/content fusion gate 0.5990039110183716 0.5990003943443298 3.476098072496825e-06 0.21980196237564087 0.6516103148460388 2.77126407623291 0.3 0.05 0.05 0.05 nan nan
6 C4 728 0.6085164835164835 0.555058605423057 206 184 338 0.6504854368932039 0.25 0.7781065088757396 0.6918120384216309 0.9515582480750588 0.5684906244277954 0.5075791281376366 0.8570581078529358 0.05379025265574453 0.8983516483516484 2.538222312927246 2.8107874393463135 1.0471160328013351 C3 plus bounded cross-time source attention and null source 0.6382710933685303 0.6382405161857605 3.0667371902382e-05 0.2080966681241989 0.6502240300178528 2.81081223487854 0.7 0.05 0.05 0.05 nan nan
7 C5 728 0.6153846153846154 0.5445555339287688 206 184 338 0.6990291262135923 0.19021739130434784 0.7958579881656804 0.6623618006706238 0.8933437077491609 0.6174448132514954 0.4918066281209682 0.8373782634735107 0.06113330221601894 0.8928571428571429 2.381938934326172 2.7620162963867188 1.212728800581531 C4 plus reliability-modulated BiGRU update 0.5485545992851257 0.5485460758209229 8.503861863573547e-06 0.13911069929599762 0.5960606336593628 2.7618556022644043 0.3 0.05 0.05 0.05 nan nan
8 C6 728 0.6263736263736264 0.5801114274083624 206 184 338 0.6941747572815534 0.29891304347826086 0.7633136094674556 0.6457235217094421 0.8814069986838935 0.6157047748565674 0.4856194350316264 0.8258335590362549 0.031823099805758544 0.885989010989011 2.30387282371521 2.774562358856201 1.1780751742536886 C5 plus optional rank-4 CP residual 0.49419134855270386 0.4941462278366089 4.511491351877339e-05 0.13551649451255798 0.5317712426185608 2.774617910385132 0.3 0.05 0.05 0.05 nan nan
9 C6_no_distance 728 0.6043956043956044 0.5144916075030433 206 184 338 0.6553398058252428 0.13043478260869565 0.8313609467455622 0.6794722676277161 0.9098390018130259 0.6045995950698853 0.5041281436421426 0.8588091731071472 0.04508186811274226 0.885989010989011 2.2977938652038574 2.808736562728882 1.36875500088095 C6 with uncertainty retained but both distance/span reliability penalties fixed to zero 0.4714658260345459 0.47136420011520386 0.00010164274135604501 0.14785172045230865 0.5429188013076782 2.8084704875946045 0.5 0.0 0.0 0.0 nan nan
10 C6_no_reconstruction 728 0.6002747252747253 0.5178357887707189 206 184 338 0.6650485436893204 0.14130434782608695 0.8106508875739645 0.6957471370697021 0.9327064671148756 0.5887437462806702 0.5024915749758389 0.8508719205856323 0.04982036520000343 0.8928571428571429 2.390105724334717 2.794447422027588 1.180521560150874 C6 trained without the auxiliary hidden-feature reconstruction loss 0.5634137392044067 0.5633965134620667 1.724438880046364e-05 0.1889886111021042 0.6063413023948669 2.794414520263672 0.3 0.05 0.05 0.05 nan nan
11 C6_pointmask 728 0.603021978021978 0.5513375927177999 206 184 338 0.587378640776699 0.2717391304347826 0.7928994082840237 0.6945403218269348 0.9553973516704227 0.5646858811378479 0.49644123517098043 0.8433728814125061 0.04921124482547844 0.8997252747252747 2.4906656742095947 2.7908360958099365 1.3967463290480098 C6 trained with independent point masking instead of contiguous spans 0.5435255169868469 0.543519139289856 6.390413091139635e-06 0.234638050198555 0.6392570734024048 2.790741205215454 0.3 0.05 0.05 0.05 nan nan
12 C7_distill 728 0.5906593406593407 0.5243846543289639 206 184 338 0.6601941747572816 0.18478260869565216 0.7692307692307693 0.6742856502532959 0.8981318409977092 0.6013374924659729 0.5035731740308492 0.8561364412307739 0.045249516246738015 0.8667582417582418 2.3018338680267334 2.8153398036956787 1.1711189639334147 C6 plus entropy/retention-weighted teacher distillation only 0.5134512186050415 0.513433575630188 1.761062230798416e-05 0.16756068170070648 0.5485631823539734 2.8153774738311768 0.3 0.05 0.05 0.05 nan nan
13 C7_group 728 0.6085164835164835 0.5116842947005743 206 184 338 0.6747572815533981 0.11413043478260869 0.8372781065088757 0.6860273480415344 0.9252878475802377 0.5911917090415955 0.5028568696567639 0.8551272749900818 0.0504213785650311 0.8928571428571429 2.445509672164917 2.7835769653320312 1.044649831831679 C6 plus smooth worst-group risk only 0.6084297895431519 0.6084184646606445 1.1323560102027841e-05 0.12680216133594513 0.6198152303695679 2.783574342727661 0.3 0.05 0.05 0.05 0.1 0.05