{ "evaluation_type": "offline paired-action metrics", "timestamp": "2026-04-17 17:27:07", "num_records": { "train": 240, "val": 80, "test": 80 }, "models": { "identity": { "train_metrics": { "mean_break_clean_l2": 0.7432431578636169, "mean_break_expert_l2": 1.4091085195541382, "mean_clean_mse": 0.3958700895309448, "mean_control_clean_l2": 0.7472800016403198, "mean_control_expert_l2": 1.4053751230239868, "mean_control_mse": 0.37037795782089233, "break_gt_control_rate_clean_ref": 0.5458333333333333, "break_gt_control_rate_expert_ref": 0.4875, "mean_clean_ref_gap": -0.004036815837025642, "mean_expert_ref_gap": 0.003733340883627534, "by_case": { "break_target_object": { "mean_break_clean_l2": 0.543735682964325, "mean_break_expert_l2": 1.5344077348709106, "mean_clean_mse": 0.45629745721817017, "mean_control_clean_l2": 0.6087575554847717, "mean_control_expert_l2": 1.5666840076446533, "mean_control_mse": 0.42383497953414917, "break_gt_control_rate_clean_ref": 0.4666666666666667, "break_gt_control_rate_expert_ref": 0.49166666666666664, "mean_clean_ref_gap": -0.06502188742160797, "mean_expert_ref_gap": -0.032276351004838943 }, "break_executability_condition": { "mean_break_clean_l2": 0.9427506923675537, "mean_break_expert_l2": 1.2838094234466553, "mean_clean_mse": 0.3354426920413971, "mean_control_clean_l2": 0.8858024477958679, "mean_control_expert_l2": 1.2440663576126099, "mean_control_mse": 0.3169209063053131, "break_gt_control_rate_clean_ref": 0.625, "break_gt_control_rate_expert_ref": 0.48333333333333334, "mean_clean_ref_gap": 0.056948255747556686, "mean_expert_ref_gap": 0.039743032306432724 } }, "num_records": 240 }, "val_metrics": { "mean_break_clean_l2": 0.6865259408950806, "mean_break_expert_l2": 1.3638770580291748, "mean_clean_mse": 0.35426950454711914, "mean_control_clean_l2": 0.6174260377883911, "mean_control_expert_l2": 1.4026581048965454, "mean_control_mse": 0.3633658289909363, "break_gt_control_rate_clean_ref": 0.575, "break_gt_control_rate_expert_ref": 0.4375, "mean_clean_ref_gap": 0.06909994781017303, "mean_expert_ref_gap": -0.03878090903162956, "by_case": { "break_target_object": { "mean_break_clean_l2": 0.6781854629516602, "mean_break_expert_l2": 1.4673945903778076, "mean_clean_mse": 0.43167275190353394, "mean_control_clean_l2": 0.5327732563018799, "mean_control_expert_l2": 1.5085182189941406, "mean_control_mse": 0.3970142602920532, "break_gt_control_rate_clean_ref": 0.575, "break_gt_control_rate_expert_ref": 0.375, "mean_clean_ref_gap": 0.14541222155094147, "mean_expert_ref_gap": -0.041123662143945694 }, "break_executability_condition": { "mean_break_clean_l2": 0.694866418838501, "mean_break_expert_l2": 1.2603596448898315, "mean_clean_mse": 0.27686622738838196, "mean_control_clean_l2": 0.7020787596702576, "mean_control_expert_l2": 1.2967978715896606, "mean_control_mse": 0.32971736788749695, "break_gt_control_rate_clean_ref": 0.575, "break_gt_control_rate_expert_ref": 0.5, "mean_clean_ref_gap": -0.007212340831756592, "mean_expert_ref_gap": -0.036438144743442535 } }, "num_records": 80 }, "test_metrics": { "mean_break_clean_l2": 0.6800279021263123, "mean_break_expert_l2": 1.6663868427276611, "mean_clean_mse": 0.39530259370803833, "mean_control_clean_l2": 0.7465093731880188, "mean_control_expert_l2": 1.351813554763794, "mean_control_mse": 0.3437272906303406, "break_gt_control_rate_clean_ref": 0.5125, "break_gt_control_rate_expert_ref": 0.5625, "mean_clean_ref_gap": -0.06648143380880356, "mean_expert_ref_gap": 0.31457358598709106, "by_case": { "break_target_object": { "mean_break_clean_l2": 0.4548130929470062, "mean_break_expert_l2": 1.7907543182373047, "mean_clean_mse": 0.4607141613960266, "mean_control_clean_l2": 0.6214543581008911, "mean_control_expert_l2": 1.5308020114898682, "mean_control_mse": 0.40925103425979614, "break_gt_control_rate_clean_ref": 0.4, "break_gt_control_rate_expert_ref": 0.55, "mean_clean_ref_gap": -0.1666412204504013, "mean_expert_ref_gap": 0.25995248556137085 }, "break_executability_condition": { "mean_break_clean_l2": 0.9052427411079407, "mean_break_expert_l2": 1.5420197248458862, "mean_clean_mse": 0.32989105582237244, "mean_control_clean_l2": 0.8715643882751465, "mean_control_expert_l2": 1.1728250980377197, "mean_control_mse": 0.27820348739624023, "break_gt_control_rate_clean_ref": 0.625, "break_gt_control_rate_expert_ref": 0.575, "mean_clean_ref_gap": 0.033678364008665085, "mean_expert_ref_gap": 0.3691946864128113 } }, "num_records": 80 } }, "action_only_adapter": { "best_val_loss_total": 1.5956528186798096, "train_metrics": { "mean_break_clean_l2": 0.36038321256637573, "mean_break_expert_l2": 1.1645487546920776, "mean_clean_mse": 0.20006108283996582, "mean_control_clean_l2": 0.2829788327217102, "mean_control_expert_l2": 1.1402603387832642, "mean_control_mse": 0.19075056910514832, "break_gt_control_rate_clean_ref": 0.5875, "break_gt_control_rate_expert_ref": 0.5291666666666667, "mean_clean_ref_gap": 0.07740439474582672, "mean_expert_ref_gap": 0.024288399145007133, "by_case": { "break_target_object": { "mean_break_clean_l2": 0.25087252259254456, "mean_break_expert_l2": 1.198777675628662, "mean_clean_mse": 0.21315599977970123, "mean_control_clean_l2": 0.22465568780899048, "mean_control_expert_l2": 1.1870900392532349, "mean_control_mse": 0.20506463944911957, "break_gt_control_rate_clean_ref": 0.5083333333333333, "break_gt_control_rate_expert_ref": 0.55, "mean_clean_ref_gap": 0.026216844096779823, "mean_expert_ref_gap": 0.011687582358717918 }, "break_executability_condition": { "mean_break_clean_l2": 0.4698939025402069, "mean_break_expert_l2": 1.1303197145462036, "mean_clean_mse": 0.1869661659002304, "mean_control_clean_l2": 0.34130197763442993, "mean_control_expert_l2": 1.0934306383132935, "mean_control_mse": 0.17643649876117706, "break_gt_control_rate_clean_ref": 0.6666666666666666, "break_gt_control_rate_expert_ref": 0.5083333333333333, "mean_clean_ref_gap": 0.12859195470809937, "mean_expert_ref_gap": 0.0368892140686512 } }, "num_records": 240 }, "val_metrics": { "mean_break_clean_l2": 0.34965047240257263, "mean_break_expert_l2": 1.1548038721084595, "mean_clean_mse": 0.1989673674106598, "mean_control_clean_l2": 0.2915192246437073, "mean_control_expert_l2": 1.1498476266860962, "mean_control_mse": 0.19426663219928741, "break_gt_control_rate_clean_ref": 0.6125, "break_gt_control_rate_expert_ref": 0.55, "mean_clean_ref_gap": 0.05813126638531685, "mean_expert_ref_gap": 0.004956287331879139, "by_case": { "break_target_object": { "mean_break_clean_l2": 0.2960914373397827, "mean_break_expert_l2": 1.1769672632217407, "mean_clean_mse": 0.2139633595943451, "mean_control_clean_l2": 0.262840211391449, "mean_control_expert_l2": 1.1790968179702759, "mean_control_mse": 0.2026643306016922, "break_gt_control_rate_clean_ref": 0.575, "break_gt_control_rate_expert_ref": 0.5, "mean_clean_ref_gap": 0.03325119987130165, "mean_expert_ref_gap": -0.002129593398422003 }, "break_executability_condition": { "mean_break_clean_l2": 0.4032095968723297, "mean_break_expert_l2": 1.1326402425765991, "mean_clean_mse": 0.1839713752269745, "mean_control_clean_l2": 0.3201982378959656, "mean_control_expert_l2": 1.1205980777740479, "mean_control_mse": 0.18586888909339905, "break_gt_control_rate_clean_ref": 0.65, "break_gt_control_rate_expert_ref": 0.6, "mean_clean_ref_gap": 0.08301133662462234, "mean_expert_ref_gap": 0.012042166665196419 } }, "num_records": 80 }, "test_metrics": { "mean_break_clean_l2": 0.3397909700870514, "mean_break_expert_l2": 1.2354549169540405, "mean_clean_mse": 0.21382936835289001, "mean_control_clean_l2": 0.3245377540588379, "mean_control_expert_l2": 1.1614665985107422, "mean_control_mse": 0.19767430424690247, "break_gt_control_rate_clean_ref": 0.5375, "break_gt_control_rate_expert_ref": 0.5125, "mean_clean_ref_gap": 0.015253203921020031, "mean_expert_ref_gap": 0.07398830354213715, "by_case": { "break_target_object": { "mean_break_clean_l2": 0.23053522408008575, "mean_break_expert_l2": 1.2631430625915527, "mean_clean_mse": 0.22174766659736633, "mean_control_clean_l2": 0.25600486993789673, "mean_control_expert_l2": 1.2138757705688477, "mean_control_mse": 0.21398238837718964, "break_gt_control_rate_clean_ref": 0.45, "break_gt_control_rate_expert_ref": 0.475, "mean_clean_ref_gap": -0.025469651445746422, "mean_expert_ref_gap": 0.04926721006631851 }, "break_executability_condition": { "mean_break_clean_l2": 0.44904670119285583, "mean_break_expert_l2": 1.2077667713165283, "mean_clean_mse": 0.2059110701084137, "mean_control_clean_l2": 0.39307063817977905, "mean_control_expert_l2": 1.1090574264526367, "mean_control_mse": 0.1813661754131317, "break_gt_control_rate_clean_ref": 0.625, "break_gt_control_rate_expert_ref": 0.55, "mean_clean_ref_gap": 0.05597605183720589, "mean_expert_ref_gap": 0.09870939701795578 } }, "num_records": 80 } }, "text_action_adapter": { "best_val_loss_total": 0.36617913842201233, "train_metrics": { "mean_break_clean_l2": 1.9662588834762573, "mean_break_expert_l2": 2.3562726974487305, "mean_clean_mse": 0.1834717094898224, "mean_control_clean_l2": 0.2672695815563202, "mean_control_expert_l2": 1.0719788074493408, "mean_control_mse": 0.1756630837917328, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.6989891529083252, "mean_expert_ref_gap": 1.2842938899993896, "by_case": { "break_target_object": { "mean_break_clean_l2": 1.7676277160644531, "mean_break_expert_l2": 2.259204864501953, "mean_clean_mse": 0.2054469883441925, "mean_control_clean_l2": 0.2036994993686676, "mean_control_expert_l2": 1.1436399221420288, "mean_control_mse": 0.19497980177402496, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.5639280080795288, "mean_expert_ref_gap": 1.1155647039413452 }, "break_executability_condition": { "mean_break_clean_l2": 2.1648900508880615, "mean_break_expert_l2": 2.4533402919769287, "mean_clean_mse": 0.16149643063545227, "mean_control_clean_l2": 0.3308396339416504, "mean_control_expert_l2": 1.0003175735473633, "mean_control_mse": 0.1563463807106018, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.8340502977371216, "mean_expert_ref_gap": 1.453023076057434 } }, "num_records": 240 }, "val_metrics": { "mean_break_clean_l2": 1.9402227401733398, "mean_break_expert_l2": 2.3244988918304443, "mean_clean_mse": 0.168838232755661, "mean_control_clean_l2": 0.23467329144477844, "mean_control_expert_l2": 1.0611259937286377, "mean_control_mse": 0.17241854965686798, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.7055492401123047, "mean_expert_ref_gap": 1.2633731365203857, "by_case": { "break_target_object": { "mean_break_clean_l2": 1.78369140625, "mean_break_expert_l2": 2.2142577171325684, "mean_clean_mse": 0.19796983897686005, "mean_control_clean_l2": 0.19418850541114807, "mean_control_expert_l2": 1.1197407245635986, "mean_control_mse": 0.18957313895225525, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.5895028114318848, "mean_expert_ref_gap": 1.0945169925689697 }, "break_executability_condition": { "mean_break_clean_l2": 2.0967538356781006, "mean_break_expert_l2": 2.4347403049468994, "mean_clean_mse": 0.13970664143562317, "mean_control_clean_l2": 0.2751580476760864, "mean_control_expert_l2": 1.0025112628936768, "mean_control_mse": 0.15526394546031952, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.8215957880020142, "mean_expert_ref_gap": 1.4322290420532227 } }, "num_records": 80 }, "test_metrics": { "mean_break_clean_l2": 1.9752228260040283, "mean_break_expert_l2": 2.467482089996338, "mean_clean_mse": 0.18618619441986084, "mean_control_clean_l2": 0.273476779460907, "mean_control_expert_l2": 1.0540869235992432, "mean_control_mse": 0.1706114113330841, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.7017459869384766, "mean_expert_ref_gap": 1.4133949279785156, "by_case": { "break_target_object": { "mean_break_clean_l2": 1.7627042531967163, "mean_break_expert_l2": 2.334259033203125, "mean_clean_mse": 0.21386170387268066, "mean_control_clean_l2": 0.2173435240983963, "mean_control_expert_l2": 1.1480493545532227, "mean_control_mse": 0.19736817479133606, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.5453606843948364, "mean_expert_ref_gap": 1.1862096786499023 }, "break_executability_condition": { "mean_break_clean_l2": 2.18774151802063, "mean_break_expert_l2": 2.6007046699523926, "mean_clean_mse": 0.1585106998682022, "mean_control_clean_l2": 0.3296099901199341, "mean_control_expert_l2": 0.9601246118545532, "mean_control_mse": 0.14385461807250977, "break_gt_control_rate_clean_ref": 1.0, "break_gt_control_rate_expert_ref": 1.0, "mean_clean_ref_gap": 1.8581314086914062, "mean_expert_ref_gap": 1.640580415725708 } }, "num_records": 80 } } } }