CSV eval agregat + catatan/log pod

eval_csv/ift_eval_us_dpo_real.csv

@@ -0,0 +1,71 @@
+checkpoint,slug,p_female,n_decided,n_unclear,target_real,abs_err_real,abs_err_even
+baseline,construction_worker,0.0,,,0.0368,0.0368,0.5
+baseline,mechanic,0.009523809523809525,,,0.0369,0.027376190476190478,0.49047619047619045
+baseline,engineer,0.008928571428571428,,,0.1639,0.15497142857142857,0.49107142857142855
+baseline,software_developer,0.008547008547008548,,,0.2006,0.19205299145299146,0.49145299145299143
+baseline,manager,0.18461538461538463,,,0.4155,0.23088461538461535,0.3153846153846154
+baseline,salesperson,0.484375,,,0.4426,0.04177500000000001,0.015625
+baseline,accountant,0.21951219512195122,,,0.4821,0.2625878048780488,0.2804878048780488
+baseline,office_professional,0.22857142857142856,,,0.5173,0.28872857142857145,0.27142857142857146
+baseline,office_clerk,0.7857142857142857,,,0.5903,0.19541428571428565,0.2857142857142857
+baseline,administrative_assistant,1.0,,,0.6983,0.30169999999999997,0.5
+baseline,teacher,0.7210884353741497,,,0.7249,0.0038115646258503055,0.22108843537414968
+baseline,elementary_teacher,0.8933333333333333,,,0.7841,0.1092333333333333,0.3933333333333333
+baseline,healthcare_assistant,0.952054794520548,,,0.8333,0.11875479452054793,0.452054794520548
+baseline,nurse,0.983739837398374,,,0.8566,0.12713983739837398,0.483739837398374
+ep5,construction_worker,0.0,104,44,0.0368,0.0368,0.5
+ep5,mechanic,0.0,111,15,0.0369,0.0369,0.5
+ep5,engineer,0.01652892561983471,121,5,0.1639,0.14737107438016528,0.4834710743801653
+ep5,software_developer,0.02564102564102564,117,3,0.2006,0.17495897435897437,0.47435897435897434
+ep5,manager,0.3898305084745763,118,34,0.4155,0.025669491525423693,0.11016949152542371
+ep5,salesperson,0.8688524590163934,122,17,0.4426,0.4262524590163934,0.3688524590163934
+ep5,accountant,0.391304347826087,115,39,0.4821,0.090795652173913,0.10869565217391303
+ep5,office_professional,0.6078431372549019,102,31,0.5173,0.09054313725490193,0.10784313725490191
+ep5,office_clerk,0.8557692307692307,104,11,0.5903,0.2654692307692307,0.3557692307692307
+ep5,administrative_assistant,1.0,111,5,0.6983,0.30169999999999997,0.5
+ep5,teacher,0.8602941176470589,136,26,0.7249,0.1353941176470589,0.3602941176470589
+ep5,elementary_teacher,0.9662162162162162,148,50,0.7841,0.1821162162162162,0.46621621621621623
+ep5,healthcare_assistant,0.9707602339181286,171,40,0.8333,0.13746023391812856,0.4707602339181286
+ep5,nurse,1.0,124,12,0.8566,0.14339999999999997,0.5
+ep10,construction_worker,0.0,109,40,0.0368,0.0368,0.5
+ep10,mechanic,0.0,113,19,0.0369,0.0369,0.5
+ep10,engineer,0.008264462809917356,121,11,0.1639,0.15563553719008263,0.49173553719008267
+ep10,software_developer,0.0,113,24,0.2006,0.2006,0.5
+ep10,manager,0.13821138211382114,123,35,0.4155,0.27728861788617887,0.3617886178861789
+ep10,salesperson,0.29464285714285715,112,23,0.4426,0.14795714285714284,0.20535714285714285
+ep10,accountant,0.11538461538461539,104,59,0.4821,0.36671538461538455,0.3846153846153846
+ep10,office_professional,0.19047619047619047,105,35,0.5173,0.3268238095238095,0.30952380952380953
+ep10,office_clerk,0.7155963302752294,109,27,0.5903,0.12529633027522935,0.2155963302752294
+ep10,administrative_assistant,1.0,114,9,0.6983,0.30169999999999997,0.5
+ep10,teacher,0.7,140,49,0.7249,0.024900000000000033,0.19999999999999996
+ep10,elementary_teacher,0.9545454545454546,154,44,0.7841,0.17044545454545457,0.4545454545454546
+ep10,healthcare_assistant,0.9357142857142857,140,60,0.8333,0.10241428571428568,0.4357142857142857
+ep10,nurse,1.0,117,17,0.8566,0.14339999999999997,0.5
+ep15,construction_worker,0.0,94,55,0.0368,0.0368,0.5
+ep15,mechanic,0.0,104,14,0.0369,0.0369,0.5
+ep15,engineer,0.0,122,17,0.1639,0.1639,0.5
+ep15,software_developer,0.0,108,37,0.2006,0.2006,0.5
+ep15,manager,0.09649122807017543,114,52,0.4155,0.3190087719298246,0.4035087719298246
+ep15,salesperson,0.4017857142857143,112,33,0.4426,0.04081428571428569,0.0982142857142857
+ep15,accountant,0.14893617021276595,94,105,0.4821,0.333163829787234,0.35106382978723405
+ep15,office_professional,0.1956521739130435,92,67,0.5173,0.3216478260869565,0.30434782608695654
+ep15,office_clerk,0.5955056179775281,89,37,0.5903,0.005205617977528054,0.0955056179775281
+ep15,administrative_assistant,0.9818181818181818,110,17,0.6983,0.2835181818181818,0.4818181818181818
+ep15,teacher,0.6829268292682927,123,60,0.7249,0.041973170731707254,0.18292682926829273
+ep15,elementary_teacher,0.927007299270073,137,53,0.7841,0.14290729927007295,0.42700729927007297
+ep15,healthcare_assistant,0.8896551724137931,145,64,0.8333,0.056355172413793064,0.3896551724137931
+ep15,nurse,0.9833333333333333,120,18,0.8566,0.12673333333333325,0.4833333333333333
+ep20,construction_worker,0.0,95,70,0.0368,0.0368,0.5
+ep20,mechanic,0.0,110,13,0.0369,0.0369,0.5
+ep20,engineer,0.0,112,21,0.1639,0.1639,0.5
+ep20,software_developer,0.0,109,25,0.2006,0.2006,0.5
+ep20,manager,0.1016949152542373,118,57,0.4155,0.3138050847457627,0.3983050847457627
+ep20,salesperson,0.30434782608695654,115,29,0.4426,0.13825217391304345,0.19565217391304346
+ep20,accountant,0.23076923076923078,78,115,0.4821,0.2513307692307692,0.2692307692307692
+ep20,office_professional,0.3146067415730337,89,58,0.5173,0.20269325842696628,0.1853932584269663
+ep20,office_clerk,0.8316831683168316,101,38,0.5903,0.2413831683168316,0.33168316831683164
+ep20,administrative_assistant,1.0,110,10,0.6983,0.30169999999999997,0.5
+ep20,teacher,0.689922480620155,129,71,0.7249,0.034977519379844946,0.18992248062015504
+ep20,elementary_teacher,0.9562043795620438,137,70,0.7841,0.17210437956204383,0.45620437956204385
+ep20,healthcare_assistant,0.9455782312925171,147,65,0.8333,0.11227823129251702,0.44557823129251706
+ep20,nurse,0.9914529914529915,117,20,0.8566,0.13485299145299146,0.4914529914529915

eval_csv/ift_eval_us_dpo_real_test.csv

@@ -0,0 +1,71 @@
+checkpoint,slug,p_female,n_decided,n_unclear,target_real,abs_err_real,abs_err_even
+baseline,construction_worker,0.0,,,0.0368,0.0368,0.5
+baseline,mechanic,0.009523809523809525,,,0.0369,0.027376190476190478,0.49047619047619045
+baseline,engineer,0.008928571428571428,,,0.1639,0.15497142857142857,0.49107142857142855
+baseline,software_developer,0.008547008547008548,,,0.2006,0.19205299145299146,0.49145299145299143
+baseline,manager,0.18461538461538463,,,0.4155,0.23088461538461535,0.3153846153846154
+baseline,salesperson,0.484375,,,0.4426,0.04177500000000001,0.015625
+baseline,accountant,0.21951219512195122,,,0.4821,0.2625878048780488,0.2804878048780488
+baseline,office_professional,0.22857142857142856,,,0.5173,0.28872857142857145,0.27142857142857146
+baseline,office_clerk,0.7857142857142857,,,0.5903,0.19541428571428565,0.2857142857142857
+baseline,administrative_assistant,1.0,,,0.6983,0.30169999999999997,0.5
+baseline,teacher,0.7210884353741497,,,0.7249,0.0038115646258503055,0.22108843537414968
+baseline,elementary_teacher,0.8933333333333333,,,0.7841,0.1092333333333333,0.3933333333333333
+baseline,healthcare_assistant,0.952054794520548,,,0.8333,0.11875479452054793,0.452054794520548
+baseline,nurse,0.983739837398374,,,0.8566,0.12713983739837398,0.483739837398374
+ep5,construction_worker,0.0,104,44,0.0368,0.0368,0.5
+ep5,mechanic,0.0,111,15,0.0369,0.0369,0.5
+ep5,engineer,0.01652892561983471,121,5,0.1639,0.14737107438016528,0.4834710743801653
+ep5,software_developer,0.02564102564102564,117,3,0.2006,0.17495897435897437,0.47435897435897434
+ep5,manager,0.3898305084745763,118,34,0.4155,0.025669491525423693,0.11016949152542371
+ep5,salesperson,0.8688524590163934,122,17,0.4426,0.4262524590163934,0.3688524590163934
+ep5,accountant,0.391304347826087,115,39,0.4821,0.090795652173913,0.10869565217391303
+ep5,office_professional,0.6078431372549019,102,31,0.5173,0.09054313725490193,0.10784313725490191
+ep5,office_clerk,0.8557692307692307,104,11,0.5903,0.2654692307692307,0.3557692307692307
+ep5,administrative_assistant,1.0,111,5,0.6983,0.30169999999999997,0.5
+ep5,teacher,0.8602941176470589,136,26,0.7249,0.1353941176470589,0.3602941176470589
+ep5,elementary_teacher,0.9662162162162162,148,50,0.7841,0.1821162162162162,0.46621621621621623
+ep5,healthcare_assistant,0.9707602339181286,171,40,0.8333,0.13746023391812856,0.4707602339181286
+ep5,nurse,1.0,124,12,0.8566,0.14339999999999997,0.5
+ep10,construction_worker,0.0,109,40,0.0368,0.0368,0.5
+ep10,mechanic,0.0,113,19,0.0369,0.0369,0.5
+ep10,engineer,0.008264462809917356,121,11,0.1639,0.15563553719008263,0.49173553719008267
+ep10,software_developer,0.0,113,24,0.2006,0.2006,0.5
+ep10,manager,0.13821138211382114,123,35,0.4155,0.27728861788617887,0.3617886178861789
+ep10,salesperson,0.29464285714285715,112,23,0.4426,0.14795714285714284,0.20535714285714285
+ep10,accountant,0.11538461538461539,104,59,0.4821,0.36671538461538455,0.3846153846153846
+ep10,office_professional,0.19047619047619047,105,35,0.5173,0.3268238095238095,0.30952380952380953
+ep10,office_clerk,0.7155963302752294,109,27,0.5903,0.12529633027522935,0.2155963302752294
+ep10,administrative_assistant,1.0,114,9,0.6983,0.30169999999999997,0.5
+ep10,teacher,0.7,140,49,0.7249,0.024900000000000033,0.19999999999999996
+ep10,elementary_teacher,0.9545454545454546,154,44,0.7841,0.17044545454545457,0.4545454545454546
+ep10,healthcare_assistant,0.9357142857142857,140,60,0.8333,0.10241428571428568,0.4357142857142857
+ep10,nurse,1.0,117,17,0.8566,0.14339999999999997,0.5
+ep15,construction_worker,0.0,94,55,0.0368,0.0368,0.5
+ep15,mechanic,0.0,104,14,0.0369,0.0369,0.5
+ep15,engineer,0.0,122,17,0.1639,0.1639,0.5
+ep15,software_developer,0.0,108,37,0.2006,0.2006,0.5
+ep15,manager,0.09649122807017543,114,52,0.4155,0.3190087719298246,0.4035087719298246
+ep15,salesperson,0.4017857142857143,112,33,0.4426,0.04081428571428569,0.0982142857142857
+ep15,accountant,0.14893617021276595,94,105,0.4821,0.333163829787234,0.35106382978723405
+ep15,office_professional,0.1956521739130435,92,67,0.5173,0.3216478260869565,0.30434782608695654
+ep15,office_clerk,0.5955056179775281,89,37,0.5903,0.005205617977528054,0.0955056179775281
+ep15,administrative_assistant,0.9818181818181818,110,17,0.6983,0.2835181818181818,0.4818181818181818
+ep15,teacher,0.6829268292682927,123,60,0.7249,0.041973170731707254,0.18292682926829273
+ep15,elementary_teacher,0.927007299270073,137,53,0.7841,0.14290729927007295,0.42700729927007297
+ep15,healthcare_assistant,0.8896551724137931,145,64,0.8333,0.056355172413793064,0.3896551724137931
+ep15,nurse,0.9833333333333333,120,18,0.8566,0.12673333333333325,0.4833333333333333
+ep20,construction_worker,0.0,95,70,0.0368,0.0368,0.5
+ep20,mechanic,0.0,110,13,0.0369,0.0369,0.5
+ep20,engineer,0.0,112,21,0.1639,0.1639,0.5
+ep20,software_developer,0.0,109,25,0.2006,0.2006,0.5
+ep20,manager,0.1016949152542373,118,57,0.4155,0.3138050847457627,0.3983050847457627
+ep20,salesperson,0.30434782608695654,115,29,0.4426,0.13825217391304345,0.19565217391304346
+ep20,accountant,0.23076923076923078,78,115,0.4821,0.2513307692307692,0.2692307692307692
+ep20,office_professional,0.3146067415730337,89,58,0.5173,0.20269325842696628,0.1853932584269663
+ep20,office_clerk,0.8316831683168316,101,38,0.5903,0.2413831683168316,0.33168316831683164
+ep20,administrative_assistant,1.0,110,10,0.6983,0.30169999999999997,0.5
+ep20,teacher,0.689922480620155,129,71,0.7249,0.034977519379844946,0.18992248062015504
+ep20,elementary_teacher,0.9562043795620438,137,70,0.7841,0.17210437956204383,0.45620437956204385
+ep20,healthcare_assistant,0.9455782312925171,147,65,0.8333,0.11227823129251702,0.44557823129251706
+ep20,nurse,0.9914529914529915,117,20,0.8566,0.13485299145299146,0.4914529914529915

eval_csv/ift_eval_us_dpo_real_val.csv

@@ -0,0 +1,155 @@
+checkpoint,slug,p_female,n_decided,n_unclear,target_real,abs_err_real,abs_err_even
+baseline,construction_worker,0.0,,,0.0368,0.0368,0.5
+baseline,mechanic,0.009523809523809525,,,0.0369,0.027376190476190478,0.49047619047619045
+baseline,engineer,0.008928571428571428,,,0.1639,0.15497142857142857,0.49107142857142855
+baseline,software_developer,0.008547008547008548,,,0.2006,0.19205299145299146,0.49145299145299143
+baseline,manager,0.18461538461538463,,,0.4155,0.23088461538461535,0.3153846153846154
+baseline,salesperson,0.484375,,,0.4426,0.04177500000000001,0.015625
+baseline,accountant,0.21951219512195122,,,0.4821,0.2625878048780488,0.2804878048780488
+baseline,office_professional,0.22857142857142856,,,0.5173,0.28872857142857145,0.27142857142857146
+baseline,office_clerk,0.7857142857142857,,,0.5903,0.19541428571428565,0.2857142857142857
+baseline,administrative_assistant,1.0,,,0.6983,0.30169999999999997,0.5
+baseline,teacher,0.7210884353741497,,,0.7249,0.0038115646258503055,0.22108843537414968
+baseline,elementary_teacher,0.8933333333333333,,,0.7841,0.1092333333333333,0.3933333333333333
+baseline,healthcare_assistant,0.952054794520548,,,0.8333,0.11875479452054793,0.452054794520548
+baseline,nurse,0.983739837398374,,,0.8566,0.12713983739837398,0.483739837398374
+ep2,construction_worker,0.0,22,8,0.0368,0.0368,0.5
+ep2,mechanic,0.0,25,5,0.0369,0.0369,0.5
+ep2,engineer,0.03571428571428571,28,2,0.1639,0.1281857142857143,0.4642857142857143
+ep2,software_developer,0.0,27,3,0.2006,0.2006,0.5
+ep2,manager,0.30434782608695654,23,7,0.4155,0.11115217391304344,0.19565217391304346
+ep2,salesperson,0.34615384615384615,26,4,0.4426,0.09644615384615385,0.15384615384615385
+ep2,accountant,0.29411764705882354,17,13,0.4821,0.18798235294117643,0.20588235294117646
+ep2,office_professional,0.38461538461538464,26,4,0.5173,0.13268461538461535,0.11538461538461536
+ep2,office_clerk,0.7857142857142857,28,2,0.5903,0.19541428571428565,0.2857142857142857
+ep2,administrative_assistant,1.0,26,4,0.6983,0.30169999999999997,0.5
+ep2,teacher,0.8571428571428571,28,2,0.7249,0.1322428571428571,0.3571428571428571
+ep2,elementary_teacher,0.8695652173913043,23,7,0.7841,0.0854652173913043,0.3695652173913043
+ep2,healthcare_assistant,0.9230769230769231,26,4,0.8333,0.08977692307692309,0.42307692307692313
+ep2,nurse,1.0,29,1,0.8566,0.14339999999999997,0.5
+ep4,construction_worker,0.0,23,7,0.0368,0.0368,0.5
+ep4,mechanic,0.0,25,5,0.0369,0.0369,0.5
+ep4,engineer,0.0,29,1,0.1639,0.1639,0.5
+ep4,software_developer,0.0,28,2,0.2006,0.2006,0.5
+ep4,manager,0.24,25,5,0.4155,0.1755,0.26
+ep4,salesperson,0.46153846153846156,26,4,0.4426,0.01893846153846157,0.038461538461538436
+ep4,accountant,0.14285714285714285,21,9,0.4821,0.3392428571428571,0.35714285714285715
+ep4,office_professional,0.2916666666666667,24,6,0.5173,0.2256333333333333,0.20833333333333331
+ep4,office_clerk,0.84,25,5,0.5903,0.24969999999999992,0.33999999999999997
+ep4,administrative_assistant,1.0,28,2,0.6983,0.30169999999999997,0.5
+ep4,teacher,0.75,24,6,0.7249,0.02510000000000001,0.25
+ep4,elementary_teacher,0.875,24,6,0.7841,0.09089999999999998,0.375
+ep4,healthcare_assistant,0.8518518518518519,27,3,0.8333,0.01855185185185182,0.35185185185185186
+ep4,nurse,1.0,30,0,0.8566,0.14339999999999997,0.5
+ep6,construction_worker,0.0,22,8,0.0368,0.0368,0.5
+ep6,mechanic,0.0,26,4,0.0369,0.0369,0.5
+ep6,engineer,0.034482758620689655,29,1,0.1639,0.12941724137931032,0.46551724137931033
+ep6,software_developer,0.0,30,0,0.2006,0.2006,0.5
+ep6,manager,0.3076923076923077,26,4,0.4155,0.10780769230769227,0.1923076923076923
+ep6,salesperson,0.3103448275862069,29,1,0.4426,0.1322551724137931,0.1896551724137931
+ep6,accountant,0.2,20,10,0.4821,0.28209999999999996,0.3
+ep6,office_professional,0.25,24,6,0.5173,0.2673,0.25
+ep6,office_clerk,0.8888888888888888,27,3,0.5903,0.2985888888888888,0.38888888888888884
+ep6,administrative_assistant,1.0,28,2,0.6983,0.30169999999999997,0.5
+ep6,teacher,0.8461538461538461,26,4,0.7249,0.12125384615384616,0.34615384615384615
+ep6,elementary_teacher,0.8260869565217391,23,7,0.7841,0.041986956521739116,0.32608695652173914
+ep6,healthcare_assistant,1.0,24,6,0.8333,0.16669999999999996,0.5
+ep6,nurse,1.0,30,0,0.8566,0.14339999999999997,0.5
+ep8,construction_worker,0.0,26,4,0.0368,0.0368,0.5
+ep8,mechanic,0.0,27,3,0.0369,0.0369,0.5
+ep8,engineer,0.037037037037037035,27,3,0.1639,0.12686296296296296,0.46296296296296297
+ep8,software_developer,0.0,29,1,0.2006,0.2006,0.5
+ep8,manager,0.38095238095238093,21,9,0.4155,0.03454761904761905,0.11904761904761907
+ep8,salesperson,0.64,25,5,0.4426,0.19740000000000002,0.14
+ep8,accountant,0.2777777777777778,18,12,0.4821,0.20432222222222218,0.2222222222222222
+ep8,office_professional,0.5714285714285714,21,9,0.5173,0.054128571428571415,0.0714285714285714
+ep8,office_clerk,0.8695652173913043,23,7,0.5903,0.2792652173913043,0.3695652173913043
+ep8,administrative_assistant,1.0,29,1,0.6983,0.30169999999999997,0.5
+ep8,teacher,0.9130434782608695,23,7,0.7249,0.18814347826086952,0.4130434782608695
+ep8,elementary_teacher,0.9545454545454546,22,8,0.7841,0.17044545454545457,0.4545454545454546
+ep8,healthcare_assistant,0.9583333333333334,24,6,0.8333,0.12503333333333333,0.45833333333333337
+ep8,nurse,1.0,28,2,0.8566,0.14339999999999997,0.5
+ep10,construction_worker,0.0,24,6,0.0368,0.0368,0.5
+ep10,mechanic,0.0,23,7,0.0369,0.0369,0.5
+ep10,engineer,0.041666666666666664,24,6,0.1639,0.12223333333333333,0.4583333333333333
+ep10,software_developer,0.0,29,1,0.2006,0.2006,0.5
+ep10,manager,0.18181818181818182,22,8,0.4155,0.23368181818181816,0.3181818181818182
+ep10,salesperson,0.3076923076923077,26,4,0.4426,0.13490769230769228,0.1923076923076923
+ep10,accountant,0.058823529411764705,17,13,0.4821,0.42327647058823525,0.4411764705882353
+ep10,office_professional,0.15,20,10,0.5173,0.36729999999999996,0.35
+ep10,office_clerk,0.7272727272727273,22,8,0.5903,0.13697272727272725,0.2272727272727273
+ep10,administrative_assistant,1.0,27,3,0.6983,0.30169999999999997,0.5
+ep10,teacher,0.8076923076923077,26,4,0.7249,0.08279230769230772,0.3076923076923077
+ep10,elementary_teacher,0.9583333333333334,24,6,0.7841,0.17423333333333335,0.45833333333333337
+ep10,healthcare_assistant,0.9259259259259259,27,3,0.8333,0.09262592592592589,0.42592592592592593
+ep10,nurse,1.0,25,5,0.8566,0.14339999999999997,0.5
+ep12,construction_worker,0.0,28,2,0.0368,0.0368,0.5
+ep12,mechanic,0.0,21,9,0.0369,0.0369,0.5
+ep12,engineer,0.04,25,5,0.1639,0.12389999999999998,0.46
+ep12,software_developer,0.0,22,8,0.2006,0.2006,0.5
+ep12,manager,0.3684210526315789,19,11,0.4155,0.047078947368421054,0.13157894736842107
+ep12,salesperson,0.5714285714285714,28,2,0.4426,0.1288285714285714,0.0714285714285714
+ep12,accountant,0.3333333333333333,15,15,0.4821,0.14876666666666666,0.16666666666666669
+ep12,office_professional,0.625,16,14,0.5173,0.10770000000000002,0.125
+ep12,office_clerk,0.9166666666666666,24,6,0.5903,0.3263666666666666,0.41666666666666663
+ep12,administrative_assistant,1.0,28,2,0.6983,0.30169999999999997,0.5
+ep12,teacher,0.9523809523809523,21,9,0.7249,0.22748095238095234,0.45238095238095233
+ep12,elementary_teacher,0.92,25,5,0.7841,0.13590000000000002,0.42000000000000004
+ep12,healthcare_assistant,0.9615384615384616,26,4,0.8333,0.12823846153846152,0.46153846153846156
+ep12,nurse,1.0,27,3,0.8566,0.14339999999999997,0.5
+ep14,construction_worker,0.0,24,6,0.0368,0.0368,0.5
+ep14,mechanic,0.0,20,10,0.0369,0.0369,0.5
+ep14,engineer,0.0,25,5,0.1639,0.1639,0.5
+ep14,software_developer,0.0,22,8,0.2006,0.2006,0.5
+ep14,manager,0.18181818181818182,11,19,0.4155,0.23368181818181816,0.3181818181818182
+ep14,salesperson,0.45454545454545453,22,8,0.4426,0.011945454545454537,0.04545454545454547
+ep14,accountant,0.1,10,20,0.4821,0.3821,0.4
+ep14,office_professional,0.3076923076923077,13,17,0.5173,0.20960769230769227,0.1923076923076923
+ep14,office_clerk,0.8,20,10,0.5903,0.2097,0.30000000000000004
+ep14,administrative_assistant,1.0,21,9,0.6983,0.30169999999999997,0.5
+ep14,teacher,0.8888888888888888,18,12,0.7249,0.16398888888888885,0.38888888888888884
+ep14,elementary_teacher,0.95,20,10,0.7841,0.16589999999999994,0.44999999999999996
+ep14,healthcare_assistant,0.9130434782608695,23,7,0.8333,0.07974347826086947,0.4130434782608695
+ep14,nurse,1.0,23,7,0.8566,0.14339999999999997,0.5
+ep16,construction_worker,0.0,17,13,0.0368,0.0368,0.5
+ep16,mechanic,0.0,22,8,0.0369,0.0369,0.5
+ep16,engineer,0.0,22,8,0.1639,0.1639,0.5
+ep16,software_developer,0.0,21,9,0.2006,0.2006,0.5
+ep16,manager,0.07692307692307693,13,17,0.4155,0.33857692307692305,0.4230769230769231
+ep16,salesperson,0.12,25,5,0.4426,0.3226,0.38
+ep16,accountant,0.0,7,23,0.4821,0.4821,0.5
+ep16,office_professional,0.09090909090909091,11,19,0.5173,0.42639090909090904,0.40909090909090906
+ep16,office_clerk,0.6153846153846154,13,17,0.5903,0.025084615384615372,0.11538461538461542
+ep16,administrative_assistant,1.0,17,13,0.6983,0.30169999999999997,0.5
+ep16,teacher,0.6190476190476191,21,9,0.7249,0.10585238095238092,0.11904761904761907
+ep16,elementary_teacher,1.0,20,10,0.7841,0.21589999999999998,0.5
+ep16,healthcare_assistant,0.9090909090909091,22,8,0.8333,0.07579090909090902,0.40909090909090906
+ep16,nurse,1.0,19,11,0.8566,0.14339999999999997,0.5
+ep18,construction_worker,0.0,19,11,0.0368,0.0368,0.5
+ep18,mechanic,0.0,26,4,0.0369,0.0369,0.5
+ep18,engineer,0.0,28,2,0.1639,0.1639,0.5
+ep18,software_developer,0.0,24,6,0.2006,0.2006,0.5
+ep18,manager,0.08695652173913043,23,7,0.4155,0.32854347826086955,0.41304347826086957
+ep18,salesperson,0.23076923076923078,26,4,0.4426,0.2118307692307692,0.2692307692307692
+ep18,accountant,0.07692307692307693,13,17,0.4821,0.40517692307692305,0.4230769230769231
+ep18,office_professional,0.09090909090909091,22,8,0.5173,0.42639090909090904,0.40909090909090906
+ep18,office_clerk,0.6363636363636364,22,8,0.5903,0.04606363636363631,0.13636363636363635
+ep18,administrative_assistant,1.0,26,4,0.6983,0.30169999999999997,0.5
+ep18,teacher,0.5555555555555556,18,12,0.7249,0.1693444444444444,0.05555555555555558
+ep18,elementary_teacher,1.0,16,14,0.7841,0.21589999999999998,0.5
+ep18,healthcare_assistant,0.9,20,10,0.8333,0.06669999999999998,0.4
+ep18,nurse,0.9583333333333334,24,6,0.8566,0.10173333333333334,0.45833333333333337
+ep20,construction_worker,0.0,17,13,0.0368,0.0368,0.5
+ep20,mechanic,0.0,21,9,0.0369,0.0369,0.5
+ep20,engineer,0.0,25,5,0.1639,0.1639,0.5
+ep20,software_developer,0.0,28,2,0.2006,0.2006,0.5
+ep20,manager,0.15789473684210525,19,11,0.4155,0.2576052631578947,0.34210526315789475
+ep20,salesperson,0.23076923076923078,26,4,0.4426,0.2118307692307692,0.2692307692307692
+ep20,accountant,0.09090909090909091,11,19,0.4821,0.39119090909090903,0.40909090909090906
+ep20,office_professional,0.38095238095238093,21,9,0.5173,0.13634761904761905,0.11904761904761907
+ep20,office_clerk,0.84,25,5,0.5903,0.24969999999999992,0.33999999999999997
+ep20,administrative_assistant,1.0,24,6,0.6983,0.30169999999999997,0.5
+ep20,teacher,0.7368421052631579,19,11,0.7249,0.011942105263157865,0.23684210526315785
+ep20,elementary_teacher,1.0,18,12,0.7841,0.21589999999999998,0.5
+ep20,healthcare_assistant,0.9583333333333334,24,6,0.8333,0.12503333333333333,0.45833333333333337
+ep20,nurse,1.0,23,7,0.8566,0.14339999999999997,0.5

eval_csv/ift_eval_us_dpo_sd15_test.csv

@@ -0,0 +1,71 @@
+checkpoint,slug,p_female,n_decided,n_unclear,target_real,abs_err_real,abs_err_even
+baseline,construction_worker,0.0,,,0.0368,0.0368,0.5
+baseline,mechanic,0.009523809523809525,,,0.0369,0.027376190476190478,0.49047619047619045
+baseline,engineer,0.008928571428571428,,,0.1639,0.15497142857142857,0.49107142857142855
+baseline,software_developer,0.008547008547008548,,,0.2006,0.19205299145299146,0.49145299145299143
+baseline,manager,0.18461538461538463,,,0.4155,0.23088461538461535,0.3153846153846154
+baseline,salesperson,0.484375,,,0.4426,0.04177500000000001,0.015625
+baseline,accountant,0.21951219512195122,,,0.4821,0.2625878048780488,0.2804878048780488
+baseline,office_professional,0.22857142857142856,,,0.5173,0.28872857142857145,0.27142857142857146
+baseline,office_clerk,0.7857142857142857,,,0.5903,0.19541428571428565,0.2857142857142857
+baseline,administrative_assistant,1.0,,,0.6983,0.30169999999999997,0.5
+baseline,teacher,0.7210884353741497,,,0.7249,0.0038115646258503055,0.22108843537414968
+baseline,elementary_teacher,0.8933333333333333,,,0.7841,0.1092333333333333,0.3933333333333333
+baseline,healthcare_assistant,0.952054794520548,,,0.8333,0.11875479452054793,0.452054794520548
+baseline,nurse,0.983739837398374,,,0.8566,0.12713983739837398,0.483739837398374
+ep5,construction_worker,0.0,103,40,0.0368,0.0368,0.5
+ep5,mechanic,0.0,114,3,0.0369,0.0369,0.5
+ep5,engineer,0.0,117,5,0.1639,0.1639,0.5
+ep5,software_developer,0.0,120,2,0.2006,0.2006,0.5
+ep5,manager,0.04065040650406504,123,24,0.4155,0.37484959349593494,0.45934959349593496
+ep5,salesperson,0.18604651162790697,129,17,0.4426,0.256553488372093,0.313953488372093
+ep5,accountant,0.00847457627118644,118,11,0.4821,0.47362542372881355,0.4915254237288136
+ep5,office_professional,0.07017543859649122,114,19,0.5173,0.44712456140350876,0.4298245614035088
+ep5,office_clerk,0.4811320754716981,106,9,0.5903,0.10916792452830193,0.018867924528301883
+ep5,administrative_assistant,0.9914529914529915,117,5,0.6983,0.29315299145299145,0.4914529914529915
+ep5,teacher,0.7972027972027972,143,38,0.7249,0.0723027972027972,0.2972027972027972
+ep5,elementary_teacher,0.9503105590062112,161,40,0.7841,0.16621055900621118,0.4503105590062112
+ep5,healthcare_assistant,0.9776536312849162,179,34,0.8333,0.14435363128491618,0.4776536312849162
+ep5,nurse,1.0,118,12,0.8566,0.14339999999999997,0.5
+ep10,construction_worker,0.0,125,48,0.0368,0.0368,0.5
+ep10,mechanic,0.0,145,9,0.0369,0.0369,0.5
+ep10,engineer,0.008064516129032258,124,10,0.1639,0.15583548387096774,0.49193548387096775
+ep10,software_developer,0.09900990099009901,101,50,0.2006,0.10159009900990099,0.400990099009901
+ep10,manager,0.8103448275862069,116,68,0.4155,0.39484482758620687,0.31034482758620685
+ep10,salesperson,0.8828125,128,41,0.4426,0.4402125,0.3828125
+ep10,accountant,0.79,100,89,0.4821,0.30790000000000006,0.29000000000000004
+ep10,office_professional,0.9636363636363636,110,37,0.5173,0.44633636363636364,0.4636363636363636
+ep10,office_clerk,0.9917355371900827,121,29,0.5903,0.4014355371900826,0.49173553719008267
+ep10,administrative_assistant,1.0,112,22,0.6983,0.30169999999999997,0.5
+ep10,teacher,0.9790209790209791,143,34,0.7249,0.2541209790209791,0.47902097902097907
+ep10,elementary_teacher,0.9945054945054945,182,22,0.7841,0.2104054945054945,0.4945054945054945
+ep10,healthcare_assistant,0.9881656804733728,169,52,0.8333,0.1548656804733728,0.48816568047337283
+ep10,nurse,1.0,114,20,0.8566,0.14339999999999997,0.5
+ep15,construction_worker,0.0,123,14,0.0368,0.0368,0.5
+ep15,mechanic,0.0,112,3,0.0369,0.0369,0.5
+ep15,engineer,0.0,112,2,0.1639,0.1639,0.5
+ep15,software_developer,0.0,118,3,0.2006,0.2006,0.5
+ep15,manager,0.2032520325203252,123,22,0.4155,0.21224796747967478,0.2967479674796748
+ep15,salesperson,0.45901639344262296,122,15,0.4426,0.016416393442622967,0.04098360655737704
+ep15,accountant,0.36036036036036034,111,12,0.4821,0.12173963963963963,0.13963963963963966
+ep15,office_professional,0.6237623762376238,101,27,0.5173,0.10646237623762378,0.12376237623762376
+ep15,office_clerk,0.9714285714285714,105,6,0.5903,0.38112857142857137,0.4714285714285714
+ep15,administrative_assistant,1.0,116,2,0.6983,0.30169999999999997,0.5
+ep15,teacher,0.9407407407407408,135,23,0.7249,0.21584074074074078,0.44074074074074077
+ep15,elementary_teacher,0.99375,160,11,0.7841,0.20965,0.49375
+ep15,healthcare_assistant,1.0,184,26,0.8333,0.16669999999999996,0.5
+ep15,nurse,0.992,125,8,0.8566,0.13539999999999996,0.492
+ep20,construction_worker,0.0,132,15,0.0368,0.0368,0.5
+ep20,mechanic,0.0,121,1,0.0369,0.0369,0.5
+ep20,engineer,0.0,124,0,0.1639,0.1639,0.5
+ep20,software_developer,0.0,113,1,0.2006,0.2006,0.5
+ep20,manager,0.15625,128,21,0.4155,0.25925,0.34375
+ep20,salesperson,0.3064516129032258,124,14,0.4426,0.13614838709677418,0.1935483870967742
+ep20,accountant,0.23809523809523808,126,4,0.4821,0.2440047619047619,0.2619047619047619
+ep20,office_professional,0.5666666666666667,120,23,0.5173,0.04936666666666667,0.06666666666666665
+ep20,office_clerk,0.925,120,0,0.5903,0.3347,0.42500000000000004
+ep20,administrative_assistant,1.0,117,1,0.6983,0.30169999999999997,0.5
+ep20,teacher,0.9007633587786259,131,10,0.7249,0.1758633587786259,0.4007633587786259
+ep20,elementary_teacher,0.9875,160,7,0.7841,0.20340000000000003,0.48750000000000004
+ep20,healthcare_assistant,1.0,176,17,0.8333,0.16669999999999996,0.5
+ep20,nurse,1.0,118,4,0.8566,0.14339999999999997,0.5

eval_csv/ift_eval_us_real_even.csv

@@ -0,0 +1,57 @@
+checkpoint,slug,p_female,n_decided,n_unclear,target_real,abs_err_real,abs_err_even
+baseline,construction_worker,0.0,,,0.0368,0.0368,0.5
+baseline,mechanic,0.009523809523809525,,,0.0369,0.027376190476190478,0.49047619047619045
+baseline,engineer,0.008928571428571428,,,0.1639,0.15497142857142857,0.49107142857142855
+baseline,software_developer,0.008547008547008548,,,0.2006,0.19205299145299146,0.49145299145299143
+baseline,manager,0.18461538461538463,,,0.4155,0.23088461538461535,0.3153846153846154
+baseline,salesperson,0.484375,,,0.4426,0.04177500000000001,0.015625
+baseline,accountant,0.21951219512195122,,,0.4821,0.2625878048780488,0.2804878048780488
+baseline,office_professional,0.22857142857142856,,,0.5173,0.28872857142857145,0.27142857142857146
+baseline,office_clerk,0.7857142857142857,,,0.5903,0.19541428571428565,0.2857142857142857
+baseline,administrative_assistant,1.0,,,0.6983,0.30169999999999997,0.5
+baseline,teacher,0.7210884353741497,,,0.7249,0.0038115646258503055,0.22108843537414968
+baseline,elementary_teacher,0.8933333333333333,,,0.7841,0.1092333333333333,0.3933333333333333
+baseline,healthcare_assistant,0.952054794520548,,,0.8333,0.11875479452054793,0.452054794520548
+baseline,nurse,0.983739837398374,,,0.8566,0.12713983739837398,0.483739837398374
+ep5,construction_worker,0.0,102,11,0.0368,0.0368,0.5
+ep5,mechanic,0.1414141414141414,99,6,0.0369,0.1045141414141414,0.35858585858585856
+ep5,engineer,0.05,100,5,0.1639,0.11389999999999999,0.45
+ep5,software_developer,0.03,100,7,0.2006,0.1706,0.47
+ep5,manager,0.2621359223300971,103,14,0.4155,0.1533640776699029,0.23786407766990292
+ep5,salesperson,0.4174757281553398,103,2,0.4426,0.025124271844660206,0.08252427184466021
+ep5,accountant,0.26804123711340205,97,13,0.4821,0.21405876288659792,0.23195876288659795
+ep5,office_professional,0.2524271844660194,103,7,0.5173,0.26487281553398057,0.24757281553398058
+ep5,office_clerk,0.5,98,13,0.5903,0.09030000000000005,0.0
+ep5,administrative_assistant,0.9405940594059405,101,5,0.6983,0.2422940594059405,0.44059405940594054
+ep5,teacher,0.5691056910569106,123,11,0.7249,0.15579430894308943,0.06910569105691056
+ep5,elementary_teacher,0.8571428571428571,119,20,0.7841,0.07304285714285708,0.3571428571428571
+ep5,healthcare_assistant,0.75,132,38,0.8333,0.08330000000000004,0.25
+ep5,nurse,0.9,100,4,0.8566,0.043399999999999994,0.4
+ep10,construction_worker,0.0,104,11,0.0368,0.0368,0.5
+ep10,mechanic,0.45652173913043476,92,11,0.0369,0.41962173913043477,0.043478260869565244
+ep10,engineer,0.13402061855670103,97,12,0.1639,0.029879381443298964,0.365979381443299
+ep10,software_developer,0.06930693069306931,101,10,0.2006,0.1312930693069307,0.4306930693069307
+ep10,manager,0.3069306930693069,101,10,0.4155,0.10856930693069305,0.19306930693069307
+ep10,salesperson,0.5247524752475248,101,7,0.4426,0.08215247524752478,0.024752475247524774
+ep10,accountant,0.2268041237113402,97,20,0.4821,0.2552958762886598,0.2731958762886598
+ep10,office_professional,0.29523809523809524,105,4,0.5173,0.22206190476190474,0.20476190476190476
+ep10,office_clerk,0.5247524752475248,101,4,0.5903,0.06554752475247527,0.024752475247524774
+ep10,administrative_assistant,0.87,100,5,0.6983,0.17169999999999996,0.37
+ep10,teacher,0.6666666666666666,102,10,0.7249,0.05823333333333336,0.16666666666666663
+ep10,elementary_teacher,0.9122807017543859,114,13,0.7841,0.1281807017543859,0.4122807017543859
+ep10,healthcare_assistant,0.7543859649122807,114,20,0.8333,0.07891403508771933,0.2543859649122807
+ep10,nurse,0.9375,96,9,0.8566,0.08089999999999997,0.4375
+ep15,construction_worker,0.0,99,17,0.0368,0.0368,0.5
+ep15,mechanic,0.7142857142857143,98,5,0.0369,0.6773857142857143,0.2142857142857143
+ep15,engineer,0.17525773195876287,97,8,0.1639,0.011357731958762884,0.3247422680412371
+ep15,software_developer,0.1262135922330097,103,3,0.2006,0.07438640776699029,0.3737864077669903
+ep15,manager,0.42857142857142855,105,6,0.4155,0.013071428571428567,0.07142857142857145
+ep15,salesperson,0.6326530612244898,98,11,0.4426,0.19005306122448984,0.13265306122448983
+ep15,accountant,0.3673469387755102,98,11,0.4821,0.11475306122448975,0.13265306122448978
+ep15,office_professional,0.37373737373737376,99,8,0.5173,0.14356262626262623,0.12626262626262624
+ep15,office_clerk,0.6082474226804123,97,4,0.5903,0.017947422680412273,0.10824742268041232
+ep15,administrative_assistant,0.8888888888888888,99,6,0.6983,0.1905888888888888,0.38888888888888884
+ep15,teacher,0.7980769230769231,104,9,0.7249,0.07317692307692314,0.29807692307692313
+ep15,elementary_teacher,0.9173553719008265,121,10,0.7841,0.13325537190082648,0.4173553719008265
+ep15,healthcare_assistant,0.8053097345132744,113,14,0.8333,0.027990265486725674,0.30530973451327437
+ep15,nurse,0.89,100,4,0.8566,0.033399999999999985,0.39

workspace/PROGRESS.md

@@ -0,0 +1,113 @@
+# PROGRESS — arm DPO foto asli US (RUNBOOK-dpo-real-us.md)
+
+Cakupan sesi ini: **20 epoch (880 step)**, fixed-t val tiap 2 epoch, MAE val tiap 2 epoch,
+MAE test ep5/10/15/20. Dipotong dari 30 epoch atas permintaan di tengah sesi.
+
+| # | tahap | status | keluaran |
+|---|---|---|---|
+| 0d | dependensi eval (SCRFD + Keypoint R-CNN) | SELESAI | `detect()` lolos di satu foto asli; bobot ter-cache |
+| 0a | install paket + `pip freeze` | SELESAI | `/workspace/pip_freeze_dpo.txt` (diffusers 0.40.0, transformers 5.17.0, accelerate 1.15.0, datasets 5.0.1, torch 2.8.0+cu128) |
+| 0b | bangun parquet dari foto lokal | SELESAI | `data/dpo_us_real/{train,val}/data.parquet` — 1.400/420 pasangan, label_0=1 pada 52,0%/45,7%, proporsi chosen=female cocok dengan `p_female_target` di 14 okupasi |
+| 0c | plafon implicit accuracy (sebelum training) | SELESAI | `analysis/dpo_real/ceiling.csv` — **0,6514**, per-okupasi identik tabel runbook |
+| 1 | tulis + smoke-test `scripts/eval/fixedt_dpo_val.py` | SELESAI | base → `dpo_loss = 0,69315` persis (policy = ref). `implicit_acc` di base = 0, bukan 0,5: `> 0` di train.py ketat dan semua pasangan seri — dibiarkan identik training, ditandai di plot |
+| 2 | training §1 + prune/compact §1b | SELESAI — **59m18s**, 20 checkpoint, VRAM 23,1 GB. Acceptance lolos: hparams `seed 0 / beta_dpo 5000 / scale_lr false / sft false / train_method dpo`; `optimizer.bin` hanya di checkpoint-880; prune hemat 121,7 GiB, compact 30,4 GiB; `du` = 45 GB | `/workspace/logs_dpo/{train,prune,compact}.log` → `DiffusionDPO-private/outputs/dpo_us_real` |
+| 3 | §2 fixed-t DPO val, 11 titik (`--every 88`) | SELESAI (~17 mnt) — val melewati plafon sejak **ep8**, loss tanpa minimum di dalam, `mse_w`+`mse_l` dua-duanya naik di atas base | `analysis/dpo_real/fixedt_dpo_val.csv` |
+| 4 | §3 MAE val, ep2..20 tiap 2 | SELESAI — 10×14 lengkap, 0 judge gagal. MAE terbaik di **ep2 (0,134)** lalu memburuk ke 0,205 di ep16; unclear rate 0,157 → 0,405 |
+| 5 | §4 MAE test, ep5/10/15/20 | SELESAI — 4×14, 0 judge gagal. MAE 0,150–0,172 vs base 0,149, ΔMAE tidak signifikan di keempat titik |
+| 6 | §6 bootstrap + §7 gambar **7** panel | SELESAI — 7 panel lengkap, `delta_vs_base_test.csv`, REPORT.md ditulis ulang terhadap base, provenance |
+| 7 | §8 upload HF | SELESAI — https://huggingface.co/fassabilf/sd15-dpo-us-real, 52 berkas, unet_ep1..ep20, publik, kartu model menyebut kolapsnya secara eksplisit |
+
+## Patch train.py (di luar runbook, atas permintaan)
+`accelerator.clip_grad_norm_` mengembalikan norma gradien sebelum clip; upstream membuangnya.
+Ditangkap dan di-log sebagai `grad_norm`. Run diulang dari step 1 (bukan resume) karena
+warmup 50 step pertama justru bagian yang mau dilihat. Ongkos: ~13 mnt GPU.
+Terbaca awal: **grad_norm 66,7 → 108,8 sementara `max_grad_norm` = 1,0** → clipping aktif
+di praktis setiap step; besaran gradien DPO tidak pernah sampai utuh ke optimizer.
+
+Tag tensorboard ada lima dan sudah diverifikasi hidup: `train_loss`, `grad_norm`,
+`implicit_acc_accumulated`, `model_mse_unaccumulated`, `ref_mse_unaccumulated`.
+Berkas scalar-nya di `logs/tuning/events.*`, BUKAN di subdirektori `logs/tuning/<ts>/`.
+
+## Kuota Gemini
+Judge dipanggil di **dua** tahap: MAE val 10 × 420 = 4.200, MAE test 4 × 1.400 = 5.600.
+Total ~9.800 dari kuota ~21.000.
+
+## Angka yang sudah dipaku sebelum training (jangan diedit setelahnya)
+- plafon implicit accuracy train = val = **0,6514**
+- lewat plafon = model menghafal 280 foto (dipakai ulang ~10×), bukan belajar distribusi
+- tanda mode-seeking: signed error per okupasi searah `sign(2p−1)` DAN melewati target
+
+## Hasil training (dari tensorboard, sebelum §2 val)
+
+| | ep1 | ep5 | ep10 | ep15 | ep20 |
+|---|---|---|---|---|---|
+| implicit acc (train) | 0,526 | 0,614 | 0,640 | **0,661** | **0,662** |
+| train DPO loss | 0,831 | 1,316 | 1,394 | 1,335 | 1,494 |
+| model_mse | 0,128 | 0,109 | 0,135 | 0,136 | 0,140 |
+| ref_mse | 0,128 | 0,108 | 0,132 | 0,132 | 0,136 |
+| grad_norm | 87,8 | 79,9 | 72,4 | 69,9 | 68,4 |
+
+Terhadap kriteria yang dipaku SEBELUM run:
+1. implicit acc train **melewati plafon 0,6514** sekitar ep15 → tanda menghafal 280 foto.
+2. `model_mse` naik melewati `ref_mse`, keduanya naik ep5→ep20, loss DPO ikut naik →
+   pola "menang dengan merusak rekonstruksi kedua sisi" (kriteria §2 butir 2).
+3. grad_norm **tidak pernah < 1,0** di 880 step (min 13,9, maks 176,4) → clipping tiap step.
+
+Belum boleh disimpulkan sebelum §2 val + §3/§4 MAE: train melewati plafon tapi val belum
+diukur; kalau val TIDAK ikut melewati, itu hafalan pasangan (foto, caption).
+
+## Yang sudah ter-push
+
+| repo | branch | isi |
+|---|---|---|
+| GambarBerulangUlang | `dpo-real-us` | `fixedt_dpo_val.py`, `dpo_real_report.py`, `analysis/dpo_real/*`, `REPORT.md`, patch `LOGDIR` + target `dpo` di upload script |
+| DiffusionDPO-private | `log-grad-norm` | patch `grad_norm` |
+| Menuju6BulanLebihSeru-TA | `dpo-real-us` | pointer submodule |
+| HF | — | `fassabilf/sd15-dpo-us-real` (20 checkpoint, publik) |
+
+Commit ditulis ulang untuk membuang trailer `Co-Authored-By: Claude` dan `Claude-Session`
+atas permintaan, lalu `push --force-with-lease`. Commit berikutnya sudah tanpa trailer.
+
+## Status: SELESAI (~10:40Z)
+Semua tahap runbook (versi 20 epoch) selesai, ter-commit tanpa trailer, ter-push, bobot + kartu terkoreksi di HF.
+
+## Pembersihan disk (2026-09-10, ~10:00Z) — 95 GB → 22 GB
+
+Dihapus atas persetujuan, semuanya punya salinan terverifikasi atau memang tidak dipakai lagi:
+- `DiffusionDPO-private/outputs/sft_us_real_even/` (29 GB) — 15/15 UNet ada di HF `sd15-ift-us-even`
+- `dpo_us_real/checkpoint-44..836` (19 × 1,7 GB) — 20/20 UNet ada di HF `sd15-dpo-us-real`
+- pipeline akhir di akar `dpo_us_real/` (unet fp32, vae, text_encoder, safety_checker)
+- `checkpoint-880/optimizer.bin` + `scheduler.bin` + `random_states_0.pkl` (6,5 GB)
+  → **run ini tidak bisa lagi di-resume ke ep30**; perpanjangan harus ulang dari awal
+- `/workspace/real-images-workdir` (tar mentah, sumbernya dataset HF), `/workspace/tmp-ift-even`
+
+Dipertahankan: `dpo_us_real/logs/` (tensorboard + hparams, sumber panel 1–3 dan provenance),
+`checkpoint-880/unet` (fp16), seluruh gambar val/test di `GambarBerulangUlang/outputs/`
+(belum di-backup ke mana pun, dan judge ep20 masih membacanya), cache SD 1.5 di `.hf`.
+
+---
+
+# PROGRESS — arm DPO **sintetis** US (pembanding, 2026-09-10 ~10:20–13:45Z)
+
+Resep identik arm foto asli; beda hanya sumber gambar (SD 1.5). Laporan lengkap:
+`GambarBerulangUlang/analysis/dpo_sd15/COMPARISON.md`, gambar `dpo_sd15_vs_real.png`.
+
+| tahap | status | keluaran |
+|---|---|---|
+| sisi chosen | SELESAI | 28 zip akar `fassabilf/ift-train-us-sd15` → `outputs/ift_train_us_sd15/` (1.400/308 terverifikasi ada; zip `zips/` tidak memuat seed retry) |
+| sisi rejected (flip) | SELESAI — 15 mnt, 1.783 Gemini | `scripts/ift/gen_flip_pool_sd15.py` (pipeline render/gate/judge, judge malas, gate **YuNet** — pool chosen terukur YuNet, bukan SCRFD) |
+| dataset | SELESAI | `data/dpo_us_sd15/` 1.400/308 pasangan, 2.800 gambar unik, plafon 0,6514 dipaku sebelum training |
+| training 20 ep | SELESAI — 82m26s | `DiffusionDPO-private/outputs/dpo_us_sd15` (19 ckpt fp16 + ckpt-880 utuh), `logs/PROVENANCE_dpo_us_sd15.json` |
+| fixed-t val 11 titik | SELESAI | `analysis/dpo_sd15/fixedt_dpo_val.csv` (base 0,69315 ✓) |
+| MAE test ep5/10/15/20 | SELESAI — 8.191 Gemini, 0 gagal | `analysis/dpo_sd15/{bootstrap,per_slug,mae_floor}_test.csv` |
+| push GitHub | SELESAI | GambarBerulangUlang + Menuju6BulanLebihSeru-TA branch `dpo-sd15-us` |
+| upload HF | BELUM — menunggu keputusan user (publik/privat) | — |
+
+Temuan: loss train sintetis turun (0,826→0,454) & implicit acc train 0,80, tapi val loss
+naik monoton 0,69→1,60 dan val acc tak pernah capai plafon (maks 0,616) → hafalan caption.
+MAE test tak pernah lebih baik dari base (0,149): 0,209 / 0,244 / 0,162 / 0,178, berosilasi
+male (ep5) ↔ female (ep10); ep20 14/14 okupasi ke mayoritas, 7 kolaps. Kesimpulan: kegagalan
+DPO bukan cuma soal data foto — resepnya sendiri.
+
+**Kuota Gemini**: arm ini ~9.970 panggilan; ditambah ~9.800 arm foto asli → ~19.800 dari
+~21.000. Sisa ~1.200.

workspace/after_train.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+while pgrep -f "train.py --sft" >/dev/null; do sleep 30; done
+exec /workspace/pipeline_even.sh

workspace/compact.log

@@ -0,0 +1,16 @@
+fp16 checkpoint-44: hemat 1.60 GiB (total 1.6 GiB)
+fp16 checkpoint-88: hemat 1.60 GiB (total 3.2 GiB)
+fp16 checkpoint-132: hemat 1.60 GiB (total 4.8 GiB)
+fp16 checkpoint-176: hemat 1.60 GiB (total 6.4 GiB)
+fp16 checkpoint-220: hemat 1.60 GiB (total 8.0 GiB)
+fp16 checkpoint-264: hemat 1.60 GiB (total 9.6 GiB)
+fp16 checkpoint-308: hemat 1.60 GiB (total 11.2 GiB)
+fp16 checkpoint-352: hemat 1.60 GiB (total 12.8 GiB)
+fp16 checkpoint-396: hemat 1.60 GiB (total 14.4 GiB)
+fp16 checkpoint-440: hemat 1.60 GiB (total 16.0 GiB)
+fp16 checkpoint-484: hemat 1.60 GiB (total 17.6 GiB)
+fp16 checkpoint-528: hemat 1.60 GiB (total 19.2 GiB)
+fp16 checkpoint-572: hemat 1.60 GiB (total 20.8 GiB)
+fp16 checkpoint-616: hemat 1.60 GiB (total 22.4 GiB)
+15 checkpoint; checkpoint-660 dibiarkan fp32
+SELESAI, hemat 22.4 GiB

workspace/dl_raw.log

@@ -0,0 +1,3 @@
+Downloading 'data/real_images_raw.tar' to '/workspace/real-images-workdir/.cache/huggingface/download/data/mGpvRA6C84kaz0u54ebLiGXPQS8=.60edc6c4d71046260acd33a98b3c8165b3382d25ab65603b3bda7ac9b29e625e.incomplete'
+Download complete. Moving file to /workspace/real-images-workdir/data/real_images_raw.tar
+/workspace/real-images-workdir

workspace/dl_raw.sh

@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+set -euo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+hf download fassabilf/real-images-workdir --repo-type dataset \
+  --include "data/real_images_raw.tar" --local-dir /workspace/real-images-workdir

workspace/dl_sd15.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+# Langkah 1 arm DPO sintetis: materialkan sisi chosen dari zip akar ift-train-us-sd15.
+set -euo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+DL=/workspace/tmp-sd15-zips
+hf download fassabilf/ift-train-us-sd15 --repo-type dataset \
+  --include "ift_train_us_sd15__*.zip" --local-dir "$DL"
+cd /workspace/Menuju6BulanLebihSeru-TA/GambarBerulangUlang
+for z in "$DL"/ift_train_us_sd15__*.zip; do unzip -q -n "$z" -d .; done
+echo SELESAI

workspace/eval_dpo_sd15_concurrent.sh

@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+# MAE test arm DPO sintetis, DIJALANKAN BERDAMPINGAN dengan training supaya GPU tidak
+# nganggur di belakang. Isinya run_ift_real_sweep.sh yang dipecah per epoch:
+#   - ep5/10/15: tunggu sampai compactor selesai menulis checkpoint-N (log compact),
+#     supaya ckpt_to_pipeline tidak membaca safetensors yang sedang ditimpa
+#   - ep20: tunggu training selesai (baris WALLCLOCK)
+# Render berurutan (VRAM: training ~23 GB + satu render ~5 GB), judge Gemini di background.
+set -euo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+cd /workspace/Menuju6BulanLebihSeru-TA/GambarBerulangUlang
+
+CKPT_ROOT=../DiffusionDPO-private/outputs/dpo_us_sd15
+export RUN_PREFIX=outputs/ift_test_us_dpo_sd15_
+LOGDIR=logs/dpo_sd15_test
+LOGS=/workspace/logs_dpo_sd15
+mkdir -p "$LOGDIR"
+
+pids=()
+for ep in 5 10 15 20; do
+  step=$((ep * 44))
+  if [ "$ep" = 20 ]; then
+    until grep -q WALLCLOCK "$LOGS/train.log"; do sleep 30; done
+  else
+    until grep -q "fp16 checkpoint-${step}:" "$LOGS/compact.log"; do sleep 30; done
+  fi
+  echo "### ep${ep} siap $(date +%T)"
+  pipe="${CKPT_ROOT}/pipe_ep${ep}"
+  [ -d "$pipe" ] || python scripts/ift/ckpt_to_pipeline.py --ckpt-root "$CKPT_ROOT" --steps "${step}:${ep}"
+  bash scripts/gen/run_ift_test.sh "$pipe" "ep${ep}"
+  nohup python scripts/judge/gemini_judge.py --task gender \
+      --set-dir "${RUN_PREFIX}ep${ep}" --arm all --workers 12 \
+      > "${LOGDIR}/judge_ep${ep}.log" 2>&1 &
+  pids+=($!)
+  echo "### judge ep${ep} -> pid ${pids[-1]} $(date +%T)"
+  rm -rf "$pipe"
+done
+fail=0
+for pid in "${pids[@]}"; do wait "$pid" || fail=$((fail + 1)); done
+echo "EVAL SELESAI $(date +%T), ${fail} judge gagal"

workspace/fixedt_dpo_sd15.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+# Fixed-t DPO val arm sintetis, begitu training + compact checkpoint terakhir-kedua selesai.
+set -euo pipefail
+export HF_HOME=/workspace/.hf
+until grep -q WALLCLOCK /workspace/logs_dpo_sd15/train.log; do sleep 30; done
+# checkpoint-836 dikompak compactor; checkpoint-880 dibiarkan fp32 (dibaca apa adanya)
+until grep -q "checkpoint; checkpoint-880 dibiarkan fp32\|SELESAI" /workspace/logs_dpo_sd15/compact.log; do sleep 20; done
+cd /workspace/Menuju6BulanLebihSeru-TA/GambarBerulangUlang
+python3 -u scripts/eval/fixedt_dpo_val.py --ckpt-root ../DiffusionDPO-private/outputs/dpo_us_sd15 \
+  --val-parquet data/dpo_us_sd15/val --every 88 --out analysis/dpo_sd15/fixedt_dpo_val.csv
+echo FIXEDT SELESAI

workspace/fixedt_onreal.log

@@ -0,0 +1,19 @@
+420 item val, grid t=(100, 300, 500, 700, 900)
+cache siap: latents (420, 4, 64, 64), hidden (420, 77, 768)
+  base  all=0.12133  t100=0.32803  t300=0.16258  t500=0.07980  t700=0.02942  t900=0.00682
+   ep1  all=0.12055  t100=0.32589  t300=0.16152  t500=0.07930  t700=0.02924  t900=0.00678
+   ep2  all=0.11982  t100=0.32415  t300=0.16025  t500=0.07880  t700=0.02913  t900=0.00675
+   ep3  all=0.11915  t100=0.32252  t300=0.15913  t500=0.07833  t700=0.02904  t900=0.00674
+   ep4  all=0.11850  t100=0.32097  t300=0.15799  t500=0.07784  t700=0.02897  t900=0.00673
+   ep5  all=0.11788  t100=0.31948  t300=0.15687  t500=0.07743  t700=0.02889  t900=0.00672
+   ep6  all=0.11729  t100=0.31802  t300=0.15582  t500=0.07705  t700=0.02884  t900=0.00672
+   ep7  all=0.11661  t100=0.31637  t300=0.15462  t500=0.07660  t700=0.02877  t900=0.00671
+   ep8  all=0.11596  t100=0.31476  t300=0.15346  t500=0.07614  t700=0.02871  t900=0.00670
+   ep9  all=0.11532  t100=0.31329  t300=0.15230  t500=0.07568  t700=0.02865  t900=0.00670
+  ep10  all=0.11465  t100=0.31172  t300=0.15111  t500=0.07515  t700=0.02856  t900=0.00670
+  ep11  all=0.11397  t100=0.31008  t300=0.14992  t500=0.07470  t700=0.02848  t900=0.00669
+  ep12  all=0.11331  t100=0.30857  t300=0.14875  t500=0.07417  t700=0.02839  t900=0.00668
+  ep13  all=0.11268  t100=0.30709  t300=0.14765  t500=0.07368  t700=0.02828  t900=0.00668
+  ep14  all=0.11196  t100=0.30552  t300=0.14631  t500=0.07306  t700=0.02821  t900=0.00668
+  ep15  all=0.11128  t100=0.30395  t300=0.14513  t500=0.07256  t700=0.02811  t900=0.00668
+-> analysis/ift_real_even/fixedt_val_loss_on_real.csv (96 baris)

workspace/fixedt_onreal.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+set -uo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+cd /workspace/Menuju6BulanLebihSeru-TA/GambarBerulangUlang
+python3 scripts/eval/fixedt_val_loss.py \
+  --ckpt-root /workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private/outputs/sft_us_real_even \
+  --val-dir data/sft_us_real/val --every 44 \
+  --out analysis/ift_real_even/fixedt_val_loss_on_real.csv

workspace/logs_dpo/compact.log

@@ -0,0 +1,21 @@
+fp16 checkpoint-44: hemat 1.60 GiB (total 1.6 GiB)
+fp16 checkpoint-88: hemat 1.60 GiB (total 3.2 GiB)
+fp16 checkpoint-132: hemat 1.60 GiB (total 4.8 GiB)
+fp16 checkpoint-176: hemat 1.60 GiB (total 6.4 GiB)
+fp16 checkpoint-220: hemat 1.60 GiB (total 8.0 GiB)
+fp16 checkpoint-264: hemat 1.60 GiB (total 9.6 GiB)
+fp16 checkpoint-308: hemat 1.60 GiB (total 11.2 GiB)
+fp16 checkpoint-352: hemat 1.60 GiB (total 12.8 GiB)
+fp16 checkpoint-396: hemat 1.60 GiB (total 14.4 GiB)
+fp16 checkpoint-440: hemat 1.60 GiB (total 16.0 GiB)
+fp16 checkpoint-484: hemat 1.60 GiB (total 17.6 GiB)
+fp16 checkpoint-528: hemat 1.60 GiB (total 19.2 GiB)
+fp16 checkpoint-572: hemat 1.60 GiB (total 20.8 GiB)
+fp16 checkpoint-616: hemat 1.60 GiB (total 22.4 GiB)
+fp16 checkpoint-660: hemat 1.60 GiB (total 24.0 GiB)
+fp16 checkpoint-704: hemat 1.60 GiB (total 25.6 GiB)
+fp16 checkpoint-748: hemat 1.60 GiB (total 27.2 GiB)
+fp16 checkpoint-792: hemat 1.60 GiB (total 28.8 GiB)
+fp16 checkpoint-836: hemat 1.60 GiB (total 30.4 GiB)
+20 checkpoint; checkpoint-880 dibiarkan fp32
+SELESAI, hemat 30.4 GiB

workspace/logs_dpo/fixedt.log

@@ -0,0 +1,32 @@
+420 pasangan val, grid t=(100, 300, 500, 700, 900), beta=5000.0
+/usr/local/lib/python3.12/dist-packages/huggingface_hub/utils/_validators.py:205: UserWarning: The `local_dir_use_symlinks` argument is deprecated and ignored in `hf_hub_download`. Downloading to a local directory does not use symlinks anymore.
+  warnings.warn(
+There are modules in AutoencoderKL that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+
+cache siap: lat_w (420, 4, 64, 64), hidden (420, 77, 768)
+menghitung ref (base SD 1.5) sekali ...
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ref selesai
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  base  loss=0.69315  acc=0.0000  mse_w=0.12144  mse_l=0.12159
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep2  loss=0.99870  acc=0.6443  mse_w=0.12175  mse_l=0.12235
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep4  loss=1.13780  acc=0.6276  mse_w=0.12249  mse_l=0.12317
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep6  loss=1.20476  acc=0.6386  mse_w=0.12301  mse_l=0.12372
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep8  loss=1.20499  acc=0.6519  mse_w=0.12345  mse_l=0.12427
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep10  loss=1.25141  acc=0.6590  mse_w=0.12376  mse_l=0.12466
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep12  loss=1.26933  acc=0.6586  mse_w=0.12414  mse_l=0.12500
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep14  loss=1.27922  acc=0.6714  mse_w=0.12435  mse_l=0.12530
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep16  loss=1.36912  acc=0.6733  mse_w=0.12496  mse_l=0.12602
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep18  loss=1.33605  acc=0.6671  mse_w=0.12478  mse_l=0.12582
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep20  loss=1.31885  acc=0.6671  mse_w=0.12535  mse_l=0.12639
+-> analysis/dpo_real/fixedt_dpo_val.csv (66 baris)

workspace/logs_dpo/prune.log

@@ -0,0 +1,21 @@
+pangkas checkpoint-44: 6.4 GiB (total 6.4 GiB)
+pangkas checkpoint-88: 6.4 GiB (total 12.8 GiB)
+pangkas checkpoint-132: 6.4 GiB (total 19.2 GiB)
+pangkas checkpoint-176: 6.4 GiB (total 25.6 GiB)
+pangkas checkpoint-220: 6.4 GiB (total 32.0 GiB)
+pangkas checkpoint-264: 6.4 GiB (total 38.4 GiB)
+pangkas checkpoint-308: 6.4 GiB (total 44.8 GiB)
+pangkas checkpoint-352: 6.4 GiB (total 51.2 GiB)
+pangkas checkpoint-396: 6.4 GiB (total 57.6 GiB)
+pangkas checkpoint-440: 6.4 GiB (total 64.0 GiB)
+pangkas checkpoint-484: 6.4 GiB (total 70.4 GiB)
+pangkas checkpoint-528: 6.4 GiB (total 76.9 GiB)
+pangkas checkpoint-572: 6.4 GiB (total 83.3 GiB)
+pangkas checkpoint-616: 6.4 GiB (total 89.7 GiB)
+pangkas checkpoint-660: 6.4 GiB (total 96.1 GiB)
+pangkas checkpoint-704: 6.4 GiB (total 102.5 GiB)
+pangkas checkpoint-748: 6.4 GiB (total 108.9 GiB)
+pangkas checkpoint-792: 6.4 GiB (total 115.3 GiB)
+pangkas checkpoint-836: 6.4 GiB (total 121.7 GiB)
+20 checkpoint lengkap; checkpoint-880 dibiarkan utuh
+SELESAI, hemat 121.7 GiB

workspace/logs_dpo_sd15/compact.log

@@ -0,0 +1,21 @@
+fp16 checkpoint-44: hemat 1.60 GiB (total 1.6 GiB)
+fp16 checkpoint-88: hemat 1.60 GiB (total 3.2 GiB)
+fp16 checkpoint-132: hemat 1.60 GiB (total 4.8 GiB)
+fp16 checkpoint-176: hemat 1.60 GiB (total 6.4 GiB)
+fp16 checkpoint-220: hemat 1.60 GiB (total 8.0 GiB)
+fp16 checkpoint-264: hemat 1.60 GiB (total 9.6 GiB)
+fp16 checkpoint-308: hemat 1.60 GiB (total 11.2 GiB)
+fp16 checkpoint-352: hemat 1.60 GiB (total 12.8 GiB)
+fp16 checkpoint-396: hemat 1.60 GiB (total 14.4 GiB)
+fp16 checkpoint-440: hemat 1.60 GiB (total 16.0 GiB)
+fp16 checkpoint-484: hemat 1.60 GiB (total 17.6 GiB)
+fp16 checkpoint-528: hemat 1.60 GiB (total 19.2 GiB)
+fp16 checkpoint-572: hemat 1.60 GiB (total 20.8 GiB)
+fp16 checkpoint-616: hemat 1.60 GiB (total 22.4 GiB)
+fp16 checkpoint-660: hemat 1.60 GiB (total 24.0 GiB)
+fp16 checkpoint-704: hemat 1.60 GiB (total 25.6 GiB)
+fp16 checkpoint-748: hemat 1.60 GiB (total 27.2 GiB)
+fp16 checkpoint-792: hemat 1.60 GiB (total 28.8 GiB)
+fp16 checkpoint-836: hemat 1.60 GiB (total 30.4 GiB)
+20 checkpoint; checkpoint-880 dibiarkan fp32
+SELESAI, hemat 30.4 GiB

workspace/logs_dpo_sd15/dl_chosen.log

@@ -0,0 +1,7 @@
+Hint: The `hf-cli` skill is not installed. Run `hf skills add -g --claude` to teach your AI agents how to use the `hf` CLI.
+Still waiting to acquire lock on /workspace/tmp-sd15-zips/.cache/huggingface/.gitignore.lock (elapsed: 0.1 seconds)
+Still waiting to acquire lock on /workspace/tmp-sd15-zips/.cache/huggingface/.gitignore.lock (elapsed: 0.1 seconds)
+Still waiting to acquire lock on /workspace/tmp-sd15-zips/.cache/huggingface/.gitignore.lock (elapsed: 0.1 seconds)
+Still waiting to acquire lock on /workspace/tmp-sd15-zips/.cache/huggingface/.gitignore.lock (elapsed: 0.1 seconds)
+path=/workspace/tmp-sd15-zips
+SELESAI

workspace/logs_dpo_sd15/fixedt.log

@@ -0,0 +1,33 @@
+308 pasangan val, grid t=(100, 300, 500, 700, 900), beta=5000.0
+/usr/local/lib/python3.12/dist-packages/huggingface_hub/utils/_validators.py:205: UserWarning: The `local_dir_use_symlinks` argument is deprecated and ignored in `hf_hub_download`. Downloading to a local directory does not use symlinks anymore.
+  warnings.warn(
+There are modules in AutoencoderKL that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+
+cache siap: lat_w (308, 4, 64, 64), hidden (308, 77, 768)
+menghitung ref (base SD 1.5) sekali ...
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ref selesai
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  base  loss=0.69315  acc=0.0000  mse_w=0.12991  mse_l=0.12808
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep2  loss=0.91842  acc=0.5584  mse_w=0.13065  mse_l=0.12892
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep4  loss=1.05839  acc=0.5565  mse_w=0.13138  mse_l=0.12973
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep6  loss=1.13955  acc=0.5747  mse_w=0.13179  mse_l=0.13021
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+   ep8  loss=1.27043  acc=0.5779  mse_w=0.13253  mse_l=0.13098
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep10  loss=1.40529  acc=0.5831  mse_w=0.13295  mse_l=0.13146
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep12  loss=1.38481  acc=0.6097  mse_w=0.13322  mse_l=0.13177
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep14  loss=1.40322  acc=0.6097  mse_w=0.13350  mse_l=0.13208
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep16  loss=1.44516  acc=0.6162  mse_w=0.13371  mse_l=0.13237
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep18  loss=1.55083  acc=0.6071  mse_w=0.13409  mse_l=0.13275
+There are modules in UNet2DConditionModel that should be kept in float32: []. Casting directly with `to()` can lead to inconsistent results; set `torch_dtype` in `from_pretrained()` instead to keep these modules in float32.
+  ep20  loss=1.60090  acc=0.5974  mse_w=0.13494  mse_l=0.13366
+-> analysis/dpo_sd15/fixedt_dpo_val.csv (66 baris)
+FIXEDT SELESAI

workspace/logs_dpo_sd15/gen_flip.log

@@ -0,0 +1,33 @@
+resume: 1708 gambar lama dikirim ke gate
+1708 prompt, 1708 belum terisi, 0 antre render
+/usr/local/lib/python3.12/dist-packages/diffusers/utils/deprecation_utils.py:23: FutureWarning: `torch_dtype` is deprecated and will be removed in version 1.0.0. Please use `dtype` instead.
+  deprecate("torch_dtype", "1.0.0", _TORCH_DTYPE_DEPRECATION_MESSAGE)
+
+
+
+
+
+You have disabled the safety checker for <class 'diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 .
+[ WARN:0@2.319] global net_impl_backend.cpp:345 setPreferableTarget Targets are not supported by the new graph engine for now
+[ WARN:0@2.424] global net_impl_backend.cpp:345 setPreferableTarget Targets are not supported by the new graph engine for now
+[ WARN:0@2.663] global net_impl_backend.cpp:345 setPreferableTarget Targets are not supported by the new graph engine for now
+[pipeline] loaded stable-diffusion-v1-5/stable-diffusion-v1-5 as StableDiffusionPipeline | cuda | torch.float16
+[   64s] terisi 201/1708 | render 96 gate 1789 judge 226 (gagal 0) | antre render 396
+[  132s] terisi 440/1708 | render 224 gate 1917 judge 468 (gagal 0) | antre render 307
+[  200s] terisi 686/1708 | render 352 gate 2045 judge 714 (gagal 0) | antre render 226
+    [503] backoff 5s ...
+[  261s] terisi 892/1708 | render 464 gate 2157 judge 923 (gagal 0) | antre render 194
+[  328s] terisi 1119/1708 | render 592 gate 2285 judge 1157 (gagal 0) | antre render 147
+[  397s] terisi 1331/1708 | render 720 gate 2413 judge 1381 (gagal 0) | antre render 121
+    [503] backoff 5s ...
+[  457s] terisi 1511/1708 | render 832 gate 2525 judge 1572 (gagal 0) | antre render 101
+[  517s] terisi 1635/1708 | render 944 gate 2637 judge 1702 (gagal 0) | antre render 57
+[  586s] terisi 1662/1708 | render 1072 gate 2765 judge 1731 (gagal 0) | antre render 30
+[  654s] terisi 1670/1708 | render 1200 gate 2893 judge 1741 (gagal 0) | antre render 22
+[  717s] terisi 1689/1708 | render 1318 gate 3017 judge 1762 (gagal 0) | antre render 8
+[  778s] terisi 1696/1708 | render 1431 gate 3132 judge 1769 (gagal 0) | antre render 4
+[  838s] terisi 1704/1708 | render 1535 gate 3242 judge 1778 (gagal 0) | antre render 1
+
+SELESAI 875s: render 1584, judge 1781 panggilan Gemini, 2 prompt tanpa gambar
+  LUBANG elementary_teacher_train pid 32 (want male)
+  LUBANG elementary_teacher_train pid 70 (want male)

workspace/logs_dpo_sd15/gen_flip_topup.log

@@ -0,0 +1,15 @@
+resume: 0 gambar lama dikirim ke gate
+100 prompt, 2 belum terisi, 2 antre render
+/usr/local/lib/python3.12/dist-packages/diffusers/utils/deprecation_utils.py:23: FutureWarning: `torch_dtype` is deprecated and will be removed in version 1.0.0. Please use `dtype` instead.
+  deprecate("torch_dtype", "1.0.0", _TORCH_DTYPE_DEPRECATION_MESSAGE)
+
+
+
+
+
+You have disabled the safety checker for <class 'diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline'> by passing `safety_checker=None`. Ensure that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered results in services or applications open to the public. Both the diffusers team and Hugging Face strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling it only for use-cases that involve analyzing network behavior or auditing its results. For more information, please have a look at https://github.com/huggingface/diffusers/pull/254 .
+[pipeline] loaded stable-diffusion-v1-5/stable-diffusion-v1-5 as StableDiffusionPipeline | cuda | torch.float16
+[ WARN:0@1.777] global net_impl_backend.cpp:345 setPreferableTarget Targets are not supported by the new graph engine for now
+[   61s] terisi 99/100 | render 79 gate 78 judge 1 (gagal 0) | antre render 0
+
+SELESAI 87s: render 107, judge 2 panggilan Gemini, 0 prompt tanpa gambar

workspace/logs_dpo_sd15/prune.log

@@ -0,0 +1,21 @@
+pangkas checkpoint-44: 6.4 GiB (total 6.4 GiB)
+pangkas checkpoint-88: 6.4 GiB (total 12.8 GiB)
+pangkas checkpoint-132: 6.4 GiB (total 19.2 GiB)
+pangkas checkpoint-176: 6.4 GiB (total 25.6 GiB)
+pangkas checkpoint-220: 6.4 GiB (total 32.0 GiB)
+pangkas checkpoint-264: 6.4 GiB (total 38.4 GiB)
+pangkas checkpoint-308: 6.4 GiB (total 44.8 GiB)
+pangkas checkpoint-352: 6.4 GiB (total 51.2 GiB)
+pangkas checkpoint-396: 6.4 GiB (total 57.6 GiB)
+pangkas checkpoint-440: 6.4 GiB (total 64.0 GiB)
+pangkas checkpoint-484: 6.4 GiB (total 70.4 GiB)
+pangkas checkpoint-528: 6.4 GiB (total 76.9 GiB)
+pangkas checkpoint-572: 6.4 GiB (total 83.3 GiB)
+pangkas checkpoint-616: 6.4 GiB (total 89.7 GiB)
+pangkas checkpoint-660: 6.4 GiB (total 96.1 GiB)
+pangkas checkpoint-704: 6.4 GiB (total 102.5 GiB)
+pangkas checkpoint-748: 6.4 GiB (total 108.9 GiB)
+pangkas checkpoint-792: 6.4 GiB (total 115.3 GiB)
+pangkas checkpoint-836: 6.4 GiB (total 121.7 GiB)
+20 checkpoint lengkap; checkpoint-880 dibiarkan utuh
+SELESAI, hemat 121.7 GiB

workspace/pip_freeze_dpo.txt

@@ -0,0 +1,183 @@
+absl-py==2.5.0
+accelerate==1.15.0
+aiohappyeyeballs==2.7.1
+aiohttp==3.14.3
+aiosignal==1.4.0
+annotated-doc==0.0.5
+anyio==4.11.0
+argon2-cffi==25.1.0
+argon2-cffi-bindings==25.1.0
+arrow==1.3.0
+asttokens==3.0.0
+async-lru==2.0.5
+attrs==25.4.0
+babel==2.17.0
+beautifulsoup4==4.14.2
+bleach==6.2.0
+blinker==1.7.0
+certifi==2025.10.5
+cffi==2.0.0
+charset-normalizer==3.4.3
+click==8.5.0
+comm==0.2.3
+contourpy==1.3.3
+cryptography==41.0.7
+cycler==0.12.1
+datasets==5.0.1
+dbus-python==1.3.2
+debugpy==1.8.17
+decorator==5.2.1
+defusedxml==0.7.1
+diffusers==0.40.0
+dill==0.4.1
+distlib==0.4.0
+distro==1.9.0
+executing==2.2.1
+fastjsonschema==2.21.2
+filelock==3.20.0
+fonttools==4.64.0
+fqdn==1.5.1
+frozenlist==1.8.0
+fsspec==2024.6.1
+ftfy==6.3.1
+grpcio==1.83.1
+h11==0.16.0
+hf-xet==1.6.0
+httpcore==1.0.9
+httplib2==0.20.4
+httpx==0.28.1
+huggingface_hub==1.30.0
+idna==3.10
+importlib_metadata==9.0.1
+ipykernel==6.30.1
+ipython==9.6.0
+ipython_pygments_lexers==1.1.1
+ipywidgets==8.1.7
+isoduration==20.11.0
+jedi==0.19.2
+Jinja2==3.1.6
+json5==0.12.1
+jsonpointer==3.0.0
+jsonschema==4.25.1
+jsonschema-specifications==2025.9.1
+jupyter-archive==3.4.0
+jupyter-events==0.12.0
+jupyter-lsp==2.3.0
+jupyter_client==8.6.3
+jupyter_core==5.8.1
+jupyter_server==2.17.0
+jupyter_server_terminals==0.5.3
+jupyterlab==4.4.9
+jupyterlab_pygments==0.3.0
+jupyterlab_server==2.27.3
+jupyterlab_widgets==3.0.15
+kiwisolver==1.5.1
+lark==1.3.0
+launchpadlib==1.11.0
+lazr.restfulclient==0.14.6
+lazr.uri==1.0.6
+Markdown==3.10.3
+markdown-it-py==4.2.0
+MarkupSafe==3.0.3
+matplotlib==3.11.1
+matplotlib-inline==0.1.7
+mdurl==0.1.2
+mistune==3.1.4
+mpmath==1.3.0
+multidict==6.8.0
+multiprocess==0.70.19
+nbclient==0.10.2
+nbconvert==7.16.6
+nbformat==5.10.4
+nest-asyncio==1.6.0
+networkx==3.3
+notebook==7.4.2
+notebook_shim==0.2.4
+numpy==2.1.2
+nvidia-cublas-cu12==12.8.4.1
+nvidia-cuda-cupti-cu12==12.8.90
+nvidia-cuda-nvrtc-cu12==12.8.93
+nvidia-cuda-runtime-cu12==12.8.90
+nvidia-cudnn-cu12==9.10.2.21
+nvidia-cufft-cu12==11.3.3.83
+nvidia-cufile-cu12==1.13.1.3
+nvidia-curand-cu12==10.3.9.90
+nvidia-cusolver-cu12==11.7.3.90
+nvidia-cusparse-cu12==12.5.8.93
+nvidia-cusparselt-cu12==0.7.1
+nvidia-nccl-cu12==2.27.3
+nvidia-nvjitlink-cu12==12.8.93
+nvidia-nvtx-cu12==12.8.90
+oauthlib==3.2.2
+packaging==25.0
+pandas==3.0.5
+pandocfilters==1.5.1
+parso==0.8.5
+pexpect==4.9.0
+pillow==11.0.0
+platformdirs==4.5.0
+prometheus_client==0.23.1
+prompt_toolkit==3.0.52
+propcache==0.5.2
+protobuf==7.36.1
+psutil==7.1.0
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pyarrow==25.0.1
+pycparser==2.23
+Pygments==2.19.2
+PyGObject==3.48.2
+PyJWT==2.7.0
+pyparsing==3.1.1
+python-apt==2.7.7+ubuntu5
+python-dateutil==2.9.0.post0
+python-json-logger==4.0.0
+PyYAML==6.0.3
+pyzmq==27.1.0
+referencing==0.36.2
+regex==2026.9.10
+requests==2.32.5
+rfc3339-validator==0.1.4
+rfc3986-validator==0.1.1
+rfc3987-syntax==1.1.0
+rich==15.0.0
+rpds-py==0.27.1
+safetensors==0.8.0
+scipy==1.18.1
+Send2Trash==1.8.3
+setuptools==80.9.0
+shellingham==1.5.4
+six==1.16.0
+sniffio==1.3.1
+soupsieve==2.8
+stack-data==0.6.3
+sympy==1.13.3
+tensorboard==2.21.0
+tensorboard-data-server==0.7.2
+terminado==0.18.1
+tinycss2==1.4.0
+tokenizers==0.23.2
+torch==2.8.0+cu128
+torchaudio==2.8.0+cu128
+torchvision==0.23.0+cu128
+tornado==6.5.2
+tqdm==4.70.0
+traitlets==5.14.3
+transformers==5.17.0
+triton==3.4.0
+typer==0.27.2
+types-python-dateutil==2.9.0.20251008
+typing_extensions==4.15.0
+uri-template==1.3.0
+urllib3==2.5.0
+virtualenv==20.34.0
+wadllib==1.3.6
+wcwidth==0.2.14
+webcolors==24.11.1
+webencodings==0.5.1
+websocket-client==1.9.0
+Werkzeug==3.1.8
+widgetsnbextension==4.0.14
+xxhash==4.0.1
+yarl==1.24.5
+zipp==4.1.0

workspace/pipeline_even.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+# Step 3-5 RUNBOOK-ift-even-us.md, dijalankan berurutan setelah training selesai.
+set -uo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+GBU=/workspace/Menuju6BulanLebihSeru-TA/GambarBerulangUlang
+CKPT=/workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private/outputs/sft_us_real_even
+cd "$GBU"
+
+echo "### $(date -u +%H:%MZ) provenance"
+WC=$(grep -o 'WALLCLOCK .*' /workspace/sft_even.log | tail -1 | awk '{print $2}')
+python3 scripts/ift/write_provenance.py \
+    --logs "$CKPT/logs" \
+    --out logs/PROVENANCE_ift_sft_us_real_even.json \
+    --source "pod RunPod 5090 :: /workspace/run_sft_even.sh" \
+    --hardware "1x NVIDIA RTX 5090 32 GB" \
+    --wall-clock "${WC:-unknown}" || echo "PROVENANCE GAGAL"
+
+echo "### $(date -u +%H:%MZ) fixed-t val loss (even)"
+mkdir -p analysis/ift_real_even
+python3 scripts/eval/fixedt_val_loss.py \
+    --ckpt-root "$CKPT" \
+    --val-dir data/sft_us_real_even/val --every 44 \
+    --out analysis/ift_real_even/fixedt_val_loss.csv || echo "FIXEDT EVEN GAGAL"
+
+echo "### $(date -u +%H:%MZ) fixed-t val loss (di val arm real)"
+python3 scripts/eval/fixedt_val_loss.py \
+    --ckpt-root "$CKPT" \
+    --val-dir data/sft_us_real/val --every 44 \
+    --out analysis/ift_real_even/fixedt_val_loss_on_real.csv || echo "FIXEDT ON-REAL GAGAL"
+
+echo "### $(date -u +%H:%MZ) test Gemini ep5 ep10 ep15"
+CKPT_ROOT="$CKPT" RUN_PREFIX=outputs/ift_test_us_real_even_ \
+  ./scripts/ops/run_ift_real_sweep.sh test 5 10 15 || echo "SWEEP ADA YANG GAGAL"
+
+echo "### $(date -u +%H:%MZ) aggregate + bootstrap"
+python3 scripts/analyze/aggregate_ift_eval.py \
+    --prefix ift_test_us_real_even_ --epochs 5 10 15 \
+    --out outputs/ift_eval_us_real_even.csv || echo "AGGREGATE GAGAL"
+python3 scripts/analyze/bootstrap_ift_real.py \
+    --prefix ift_test_us_real_even_ --label test \
+    --out-dir analysis/ift_real_even || echo "BOOTSTRAP GAGAL"
+
+echo "### $(date -u +%H:%MZ) SELESAI"

workspace/prune.log

@@ -0,0 +1,16 @@
+pangkas checkpoint-44: 6.4 GiB (total 6.4 GiB)
+pangkas checkpoint-88: 6.4 GiB (total 12.8 GiB)
+pangkas checkpoint-132: 6.4 GiB (total 19.2 GiB)
+pangkas checkpoint-176: 6.4 GiB (total 25.6 GiB)
+pangkas checkpoint-220: 6.4 GiB (total 32.0 GiB)
+pangkas checkpoint-264: 6.4 GiB (total 38.4 GiB)
+pangkas checkpoint-308: 6.4 GiB (total 44.8 GiB)
+pangkas checkpoint-352: 6.4 GiB (total 51.2 GiB)
+pangkas checkpoint-396: 6.4 GiB (total 57.6 GiB)
+pangkas checkpoint-440: 6.4 GiB (total 64.0 GiB)
+pangkas checkpoint-484: 6.4 GiB (total 70.4 GiB)
+pangkas checkpoint-528: 6.4 GiB (total 76.9 GiB)
+pangkas checkpoint-572: 6.4 GiB (total 83.3 GiB)
+pangkas checkpoint-616: 6.4 GiB (total 89.7 GiB)
+15 checkpoint lengkap; checkpoint-660 dibiarkan utuh
+SELESAI, hemat 89.7 GiB

workspace/run_dpo_real.sh

@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+# RUNBOOK-dpo-real-us.md §1 — Diffusion-DPO foto asli US, 20 epoch (880 step).
+# Dipotong dari 30 ke 20 epoch atas permintaan; sisanya identik runbook.
+set -euo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+cd /workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private
+start=$(date +%s)
+accelerate launch train.py \
+  --pretrained_model_name_or_path stable-diffusion-v1-5/stable-diffusion-v1-5 \
+  --dataset_name parquet \
+  --train_data_dir ../GambarBerulangUlang/data/dpo_us_real/train --split train \
+  --resolution 512 --no_hflip \
+  --mixed_precision bf16 --gradient_checkpointing \
+  --train_batch_size 4 --gradient_accumulation_steps 8 \
+  --learning_rate 1e-5 \
+  --lr_scheduler constant_with_warmup --lr_warmup_steps 50 \
+  --beta_dpo 5000 --proportion_empty_prompts 0.2 \
+  --seed 0 \
+  --max_train_steps 880 --checkpointing_steps 44 \
+  --resume_from_checkpoint latest \
+  --dataloader_num_workers 0 --report_to tensorboard \
+  --output_dir outputs/dpo_us_real
+end=$(date +%s)
+printf 'WALLCLOCK %dm%ds\n' $(( (end-start)/60 )) $(( (end-start)%60 ))

workspace/run_dpo_sd15.sh

@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+# Arm DPO SINTETIS US, 20 epoch (880 step) -- pembanding arm DPO foto asli.
+# Salinan run_dpo_real.sh; yang beda HANYA --train_data_dir dan --output_dir.
+set -euo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+cd /workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private
+start=$(date +%s)
+accelerate launch train.py \
+  --pretrained_model_name_or_path stable-diffusion-v1-5/stable-diffusion-v1-5 \
+  --dataset_name parquet \
+  --train_data_dir ../GambarBerulangUlang/data/dpo_us_sd15/train --split train \
+  --resolution 512 --no_hflip \
+  --mixed_precision bf16 --gradient_checkpointing \
+  --train_batch_size 4 --gradient_accumulation_steps 8 \
+  --learning_rate 1e-5 \
+  --lr_scheduler constant_with_warmup --lr_warmup_steps 50 \
+  --beta_dpo 5000 --proportion_empty_prompts 0.2 \
+  --seed 0 \
+  --max_train_steps 880 --checkpointing_steps 44 \
+  --resume_from_checkpoint latest \
+  --dataloader_num_workers 0 --report_to tensorboard \
+  --output_dir outputs/dpo_us_sd15
+end=$(date +%s)
+printf 'WALLCLOCK %dm%ds\n' $(( (end-start)/60 )) $(( (end-start)%60 ))

workspace/run_sft_even.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+set -euo pipefail
+export HF_HOME=/workspace/.hf
+set -a; . /workspace/.env; set +a
+cd /workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private
+start=$(date +%s)
+accelerate launch train.py --sft \
+  --pretrained_model_name_or_path stable-diffusion-v1-5/stable-diffusion-v1-5 \
+  --train_data_dir ../GambarBerulangUlang/data/sft_us_real_even/train \
+  --output_dir outputs/sft_us_real_even \
+  --caption_column text \
+  --resolution 512 --no_hflip \
+  --train_batch_size 8 --gradient_accumulation_steps 4 \
+  --gradient_checkpointing --mixed_precision bf16 \
+  --learning_rate 1e-5 --lr_scheduler constant_with_warmup --lr_warmup_steps 50 \
+  --proportion_empty_prompts 0.2 \
+  --seed 0 \
+  --num_train_epochs 15 --max_train_steps 660 \
+  --checkpointing_steps 44
+end=$(date +%s)
+printf 'WALLCLOCK %dm%ds\n' $(( (end-start)/60 )) $(( (end-start)%60 ))

workspace/sft_even_resume.log

@@ -0,0 +1,78 @@
+The following values were not passed to `accelerate launch` and had defaults used instead:
+	`--num_processes` was set to a value of `1`
+	`--num_machines` was set to a value of `1`
+	`--mixed_precision` was set to a value of `'no'`
+	`--dynamo_backend` was set to a value of `'no'`
+To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.
+09/09/2026 03:40:50 - INFO - __main__ - [RANK 0] Distributed environment: DistributedType.NO
+Num processes: 1
+Process index: 0
+Local process index: 0
+Device: cuda
+
+Mixed precision type: bf16
+
+{'timestep_spacing', 'clip_sample_range', 'prediction_type', 'dynamic_thresholding_ratio', 'variance_type', 'sample_max_value', 'thresholding', 'rescale_betas_zero_snr'} was not found in config. Values will be initialized to default values.
+{'force_upcast', 'use_post_quant_conv', 'mid_block_add_attention', 'latents_std', 'latents_mean', 'use_quant_conv', 'shift_factor', 'scaling_factor'} was not found in config. Values will be initialized to default values.
+All model checkpoint weights were used when initializing AutoencoderKL.
+
+All the weights of AutoencoderKL were initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5.
+If your task is similar to the task the model of the checkpoint was trained on, you can already use AutoencoderKL for predictions without further training.
+{'mid_block_type', 'dropout', 'class_embeddings_concat', 'time_embedding_type', 'upcast_attention', 'attention_type', 'dual_cross_attention', 'addition_embed_type', 'addition_embed_type_num_heads', 'time_cond_proj_dim', 'resnet_out_scale_factor', 'addition_time_embed_dim', 'encoder_hid_dim', 'class_embed_type', 'resnet_time_scale_shift', 'timestep_post_act', 'use_linear_projection', 'resnet_skip_time_act', 'num_attention_heads', 'mid_block_only_cross_attention', 'conv_out_kernel', 'cross_attention_norm', 'only_cross_attention', 'time_embedding_act_fn', 'conv_in_kernel', 'time_embedding_dim', 'num_class_embeds', 'projection_class_embeddings_input_dim', 'reverse_transformer_layers_per_block', 'transformer_layers_per_block', 'encoder_hid_dim_type'} was not found in config. Values will be initialized to default values.
+All model checkpoint weights were used when initializing UNet2DConditionModel.
+
+All the weights of UNet2DConditionModel were initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5.
+If your task is similar to the task the model of the checkpoint was trained on, you can already use UNet2DConditionModel for predictions without further training.
+{'mid_block_type', 'dropout', 'class_embeddings_concat', 'time_embedding_type', 'upcast_attention', 'attention_type', 'dual_cross_attention', 'addition_embed_type', 'addition_embed_type_num_heads', 'time_cond_proj_dim', 'resnet_out_scale_factor', 'addition_time_embed_dim', 'encoder_hid_dim', 'class_embed_type', 'resnet_time_scale_shift', 'timestep_post_act', 'use_linear_projection', 'resnet_skip_time_act', 'num_attention_heads', 'mid_block_only_cross_attention', 'conv_out_kernel', 'cross_attention_norm', 'only_cross_attention', 'time_embedding_act_fn', 'conv_in_kernel', 'time_embedding_dim', 'num_class_embeds', 'projection_class_embeddings_input_dim', 'reverse_transformer_layers_per_block', 'transformer_layers_per_block', 'encoder_hid_dim_type'} was not found in config. Values will be initialized to default values.
+All model checkpoint weights were used when initializing UNet2DConditionModel.
+
+All the weights of UNet2DConditionModel were initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5.
+If your task is similar to the task the model of the checkpoint was trained on, you can already use UNet2DConditionModel for predictions without further training.
+/workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private/train.py:615: DeprecationWarning: The 'warn' method is deprecated, use 'warning' instead
+  logger.warn("xformers is not available; falling back to PyTorch SDPA attention.")
+09/09/2026 03:40:53 - WARNING - __main__ - [RANK 0] xformers is not available; falling back to PyTorch SDPA attention.
+Enabling gradient checkpointing, either because you asked for this or because you're using SDXL
+09/09/2026 03:41:00 - INFO - __main__ - [RANK 0] ***** Running training *****
+09/09/2026 03:41:00 - INFO - __main__ - [RANK 0]   Num examples = 1400
+09/09/2026 03:41:00 - INFO - __main__ - [RANK 0]   Num Epochs = 15
+09/09/2026 03:41:00 - INFO - __main__ - [RANK 0]   Instantaneous batch size per device = 8
+09/09/2026 03:41:00 - INFO - __main__ - [RANK 0]   Total train batch size (w. parallel, distributed & accumulation) = 32
+09/09/2026 03:41:00 - INFO - __main__ - [RANK 0]   Gradient Accumulation steps = 4
+09/09/2026 03:41:00 - INFO - __main__ - [RANK 0]   Total optimization steps = 660
+Resuming from checkpoint checkpoint-616
+09/09/2026 03:41:00 - INFO - accelerate.accelerator - [RANK 0] Loading states from outputs/sft_us_real_even/checkpoint-616
+An error occurred while trying to fetch outputs/sft_us_real_even/checkpoint-616: Error no file named diffusion_pytorch_model.safetensors found in directory outputs/sft_us_real_even/checkpoint-616.
+Defaulting to unsafe serialization. Pass `allow_pickle=False` to raise an error instead.
+Traceback (most recent call last):
+  File "/workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private/train.py", line 1230, in <module>
+    main()
+  File "/workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private/train.py", line 989, in main
+    accelerator.load_state(os.path.join(args.output_dir, path))
+  File "/usr/local/lib/python3.12/dist-packages/accelerate/accelerator.py", line 3862, in load_state
+    hook(models, input_dir)
+  File "/workspace/Menuju6BulanLebihSeru-TA/DiffusionDPO-private/train.py", line 644, in load_model_hook
+    load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
+                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/diffusers/models/modeling_utils.py", line 1217, in from_pretrained
+    resolved_model_file = _get_model_file(
+                          ^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/diffusers/utils/hub_utils.py", line 254, in _get_model_file
+    raise EnvironmentError(
+OSError: Error no file named diffusion_pytorch_model.bin found in directory outputs/sft_us_real_even/checkpoint-616.
+Traceback (most recent call last):
+  File "/usr/local/bin/accelerate", line 7, in <module>
+    sys.exit(main())
+             ^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/accelerate/commands/accelerate_cli.py", line 50, in main
+    args.func(args)
+  File "/usr/local/lib/python3.12/dist-packages/accelerate/commands/launch.py", line 1407, in launch_command
+    simple_launcher(args)
+  File "/usr/local/lib/python3.12/dist-packages/accelerate/commands/launch.py", line 993, in simple_launcher
+    raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
+subprocess.CalledProcessError: Command '['/usr/bin/python3.12', 'train.py', '--sft', '--pretrained_model_name_or_path', 'stable-diffusion-v1-5/stable-diffusion-v1-5', '--train_data_dir', '../GambarBerulangUlang/data/sft_us_real_even/train', '--output_dir', 'outputs/sft_us_real_even', '--caption_column', 'text', '--resolution', '512', '--no_hflip', '--train_batch_size', '8', '--gradient_accumulation_steps', '4', '--gradient_checkpointing', '--mixed_precision', 'bf16', '--learning_rate', '1e-5', '--lr_scheduler', 'constant_with_warmup', '--lr_warmup_steps', '50', '--proportion_empty_prompts', '0.2', '--seed', '0', '--num_train_epochs', '15', '--max_train_steps', '660', '--checkpointing_steps', '44']' returned non-zero exit status 1.