diff options
| author | YurenHao0426 <blackhao0426@gmail.com> | 2026-06-13 12:35:36 -0500 |
|---|---|---|
| committer | YurenHao0426 <blackhao0426@gmail.com> | 2026-06-13 12:35:36 -0500 |
| commit | 66e0d8b9fd4d0f7a2231d689c055e26fdf1cf04a (patch) | |
| tree | c29cba61124018755a19b02c9d33e3ad5f2e05cc /research/flossing/late_perturb_robustness/trm_multi4_best_step35805_n3000_k8_late.summary.csv | |
Curated export for clone-and-run Maze training (2x A6000) + diagnostics.
trm/hrm pretrain.py carry trajectory-augmentation code (backward-compatible).
Heavy artifacts (checkpoints/wandb/npz) gitignored; see PROVENANCE.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Diffstat (limited to 'research/flossing/late_perturb_robustness/trm_multi4_best_step35805_n3000_k8_late.summary.csv')
| -rw-r--r-- | research/flossing/late_perturb_robustness/trm_multi4_best_step35805_n3000_k8_late.summary.csv | 31 |
1 files changed, 31 insertions, 0 deletions
diff --git a/research/flossing/late_perturb_robustness/trm_multi4_best_step35805_n3000_k8_late.summary.csv b/research/flossing/late_perturb_robustness/trm_multi4_best_step35805_n3000_k8_late.summary.csv new file mode 100644 index 0000000..01a8526 --- /dev/null +++ b/research/flossing/late_perturb_robustness/trm_multi4_best_step35805_n3000_k8_late.summary.csv @@ -0,0 +1,31 @@ +label,perturb_after,sigma,n_samples,rollouts,ckpt_root,ckpt_name,perturb,noise_distribution,mean_rollout_exact,mean_rollout_token_acc,pass_at_k,all_k,correct_count_mean,correct_count_std,zero_frac,full_frac,clean_acc,retain_mean_on_clean_success,allK_on_clean_success,rescue_mean_on_clean_fail,passK_on_clean_fail
+trm_multi4_best,0,0.0,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.959773600101471,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,0,0.001,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8957083225250244,0.9601898789405823,0.9583333134651184,0.8119999766349792,7.165666580200195,2.081879138946533,0.0416666679084301,0.8119999766349792,0.8939999938011169,0.9684004187583923,0.9030573964118958,0.28262579441070557,0.6194968819618225
+trm_multi4_best,0,0.003,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8957916498184204,0.9602490067481995,0.9606666564941406,0.8130000233650208,7.166333198547363,2.062037706375122,0.03933333232998848,0.8130000233650208,0.8939999938011169,0.9666293859481812,0.9038031101226807,0.2983490526676178,0.6352201104164124
+trm_multi4_best,0,0.01,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8947083353996277,0.9597618579864502,0.9599999785423279,0.8063333630561829,7.1576666831970215,2.0646567344665527,0.03999999910593033,0.8063333630561829,0.8939999938011169,0.9667226076126099,0.8963460326194763,0.28734275698661804,0.6320754885673523
+trm_multi4_best,0,0.03,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8953750133514404,0.9600813984870911,0.9603333473205566,0.8009999990463257,7.163000106811523,2.049495220184326,0.03966666758060455,0.8009999990463257,0.8939999938011169,0.9666759967803955,0.8918717503547668,0.29402515292167664,0.6446540951728821
+trm_multi4_best,0,0.1,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8950416445732117,0.9601003527641296,0.9629999995231628,0.7983333468437195,7.160333156585693,2.0425376892089844,0.03700000047683716,0.7983333468437195,0.8939999938011169,0.9635066986083984,0.8851603269577026,0.3176100552082062,0.6666666865348816
+trm_multi4_best,4,0.0,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.959773600101471,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,4,0.001,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8946250081062317,0.9599500894546509,0.9473333358764648,0.8433333039283752,7.1570000648498535,2.1906659603118896,0.052666667848825455,0.8433333039283752,0.8939999938011169,0.9753448963165283,0.9392244815826416,0.21383647620677948,0.5251572132110596
+trm_multi4_best,4,0.003,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8962500095367432,0.9606265425682068,0.9523333311080933,0.843999981880188,7.170000076293945,2.1701996326446533,0.04766666516661644,0.843999981880188,0.8939999938011169,0.976416826248169,0.9410887360572815,0.22012577950954437,0.5691823959350586
+trm_multi4_best,4,0.01,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8963750004768372,0.9604902863502502,0.9483333230018616,0.8446666598320007,7.171000003814697,2.171119213104248,0.05166666582226753,0.8446666598320007,0.8939999938011169,0.9760906100273132,0.9410887360572815,0.22405660152435303,0.5314465165138245
+trm_multi4_best,4,0.03,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8968333601951599,0.9606810808181763,0.9506666660308838,0.8403333425521851,7.174666881561279,2.1587400436401367,0.04933333396911621,0.8403333425521851,0.8939999938011169,0.9759973883628845,0.9377330541610718,0.2291666716337204,0.544025182723999
+trm_multi4_best,4,0.1,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8944583535194397,0.9597752094268799,0.9526666402816772,0.8326666951179504,7.155666828155518,2.1656641960144043,0.047333333641290665,0.8326666951179504,0.8939999938011169,0.9737602472305298,0.9284116625785828,0.22562892735004425,0.5628930926322937
+trm_multi4_best,8,0.0,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.959773600101471,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,8,0.001,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8962083458900452,0.9605055451393127,0.9399999976158142,0.8696666955947876,7.169666767120361,2.2701423168182373,0.05999999865889549,0.8696666955947876,0.8939999938011169,0.9840604066848755,0.9709172248840332,0.15526729822158813,0.45597484707832336
+trm_multi4_best,8,0.003,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8970000147819519,0.9608615636825562,0.9416666626930237,0.8696666955947876,7.176000118255615,2.254556179046631,0.05833333358168602,0.8696666955947876,0.8939999938011169,0.9844798445701599,0.9709172248840332,0.1591981202363968,0.46855345368385315
+trm_multi4_best,8,0.01,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8955416679382324,0.9602766633033752,0.9399999976158142,0.8673333525657654,7.164333343505859,2.2744951248168945,0.05999999865889549,0.8673333525657654,0.8939999938011169,0.9835943579673767,0.9690529704093933,0.1529088020324707,0.45597484707832336
+trm_multi4_best,8,0.03,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8949999809265137,0.9600179195404053,0.9436666369438171,0.8650000095367432,7.159999847412109,2.2588493824005127,0.056333333253860474,0.8650000095367432,0.8939999938011169,0.9836875200271606,0.9668158292770386,0.14701257646083832,0.47484275698661804
+trm_multi4_best,8,0.1,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8958333134651184,0.9602567553520203,0.9449999928474426,0.8643333315849304,7.166666507720947,2.251863479614258,0.054999999701976776,0.8643333315849304,0.8939999938011169,0.9831282496452332,0.9656972289085388,0.1595911979675293,0.49685534834861755
+trm_multi4_best,12,0.0,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.959773600101471,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,12,0.001,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8956666588783264,0.9604059457778931,0.9129999876022339,0.887333333492279,7.165333271026611,2.3898391723632812,0.08699999749660492,0.887333333492279,0.8939999938011169,0.9954324960708618,0.9925428628921509,0.05424528196454048,0.19182389974594116
+trm_multi4_best,12,0.003,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8966249823570251,0.9608678221702576,0.9193333387374878,0.887333333492279,7.172999858856201,2.3672215938568115,0.0806666687130928,0.887333333492279,0.8939999938011169,0.995712161064148,0.9925428628921509,0.06092767417430878,0.24528302252292633
+trm_multi4_best,12,0.01,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8974583148956299,0.9608935713768005,0.9213333129882812,0.887666642665863,7.179666519165039,2.3568453788757324,0.07866666465997696,0.887666642665863,0.8939999938011169,0.9956655502319336,0.9925428628921509,0.069182388484478,0.26729559898376465
+trm_multi4_best,12,0.03,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8966249823570251,0.9606748819351196,0.9223333597183228,0.8870000243186951,7.172999858856201,2.3598878383636475,0.07766667008399963,0.8870000243186951,0.8939999938011169,0.9950596690177917,0.9921700358390808,0.06643081456422806,0.276729553937912
+trm_multi4_best,12,0.1,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8972083330154419,0.9610205292701721,0.9259999990463257,0.8849999904632568,7.177666664123535,2.338397264480591,0.07400000095367432,0.8849999904632568,0.8939999938011169,0.9935682415962219,0.9899328947067261,0.08451257646083832,0.31446540355682373
+trm_multi4_best,15,0.0,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.959773600101471,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,15,0.001,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.9598419666290283,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,15,0.003,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.9597874879837036,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,15,0.01,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8939999938011169,0.9598518013954163,0.8939999938011169,0.8939999938011169,7.1519999504089355,2.462700843811035,0.10599999874830246,0.8939999938011169,0.8939999938011169,1.0,1.0,0.0,0.0
+trm_multi4_best,15,0.03,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8941249847412109,0.9598509073257446,0.8946666717529297,0.8939999938011169,7.1529998779296875,2.4601335525512695,0.10533333569765091,0.8939999938011169,0.8939999938011169,1.0,1.0,0.001179245300590992,0.006289307959377766
+trm_multi4_best,15,0.1,3000,8,/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_multi4_loguniform_repro,step_35805,both,gaussian,0.8942499756813049,0.9599588513374329,0.8956666588783264,0.893666684627533,7.1539998054504395,2.456070899963379,0.10433333367109299,0.893666684627533,0.8939999938011169,0.9999067783355713,0.9996271729469299,0.003144653979688883,0.015723271295428276
|
