{ "objective": -0.5112698517216128, "paired_flip": { "pairs": 80, "flip_rate": 0.675, "both_correct_rate": 0.625, "invariant_pairs": 32, "invariance_rate": 0.96875, "invariant_both_correct_rate": 0.59375 }, "unknowable": null, "clean": { "n": 1264, "nll": 0.4989297577819544, "acc": 0.8243670886075949, "ece": 0.05704976909178335, "brier": 0.25535485232026645, "mean_conf": 0.8785235750178558, "confident_error_rate": 0.040348101265822785, "coverage_at_0_9": 0.6566455696202531, "accuracy_at_0_9": 0.9385542168674699, "coverage_at_5pct_error": 0.6337025316455697, "coverage_at_1pct_error": 0.18908227848101267, "aurc": 0.057098658886616284, "error_rate_at_0_9": 0.06144578313253012, "confidence_bias": 0.05415648641026094, "top_bins": { "0.9": { "n": 830, "errors": 51, "error_rate": 0.06144578313253012 }, "0.95": { "n": 765, "errors": 32, "error_rate": 0.04183006535947712 }, "0.99": { "n": 616, "errors": 21, "error_rate": 0.03409090909090909 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9651898734177216, "confidence_cutoff": 0.9884021799746481 }, "0.8": { "coverage": 0.8006329113924051, "accuracy": 0.8913043478260869, "confidence_cutoff": 0.7007823799483796 } }, "score_mae": 0.44288336559433733, "ranked_probability_score": 0.07198397780180195 }, "tasks": { "agnews": { "n": 80, "nll": 0.7033810609329125, "acc": 0.85, "ece": 0.12788153266590807, "brier": 0.2690347009363626, "mean_conf": 0.9626452504559462, "confident_error_rate": 0.1125, "coverage_at_0_9": 0.9, "accuracy_at_0_9": 0.875, "coverage_at_5pct_error": 0.5, "coverage_at_1pct_error": 0.175, "aurc": 0.06388995806168134, "error_rate_at_0_9": 0.125, "confidence_bias": 0.1126452504559462, "top_bins": { "0.9": { "n": 72, "errors": 9, "error_rate": 0.125 }, "0.95": { "n": 66, "errors": 6, "error_rate": 0.09090909090909091 }, "0.99": { "n": 61, "errors": 5, "error_rate": 0.08196721311475409 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.95, "confidence_cutoff": 0.9983029203561053 }, "0.8": { "coverage": 0.8, "accuracy": 0.921875, "confidence_cutoff": 0.9797347727687692 } } }, "agnews_yn": { "n": 160, "nll": 0.218266217089547, "acc": 0.925, "ece": 0.053822844561075917, "brier": 0.12359771772019439, "mean_conf": 0.9648416375603937, "confident_error_rate": 0.03125, "coverage_at_0_9": 0.9, "accuracy_at_0_9": 0.9652777777777778, "coverage_at_5pct_error": 0.94375, "coverage_at_1pct_error": 0.5, "aurc": 0.013180196924715198, "error_rate_at_0_9": 0.034722222222222224, "confidence_bias": 0.039841637560393606, "top_bins": { "0.9": { "n": 144, "errors": 5, "error_rate": 0.034722222222222224 }, "0.95": { "n": 132, "errors": 3, "error_rate": 0.022727272727272728 }, "0.99": { "n": 98, "errors": 1, "error_rate": 0.01020408163265306 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9943423697163807 }, "0.8": { "coverage": 0.8, "accuracy": 0.9765625, "confidence_cutoff": 0.962791422719902 } } }, "amazon": { "n": 80, "nll": 0.9101254335951751, "acc": 0.65, "ece": 0.09852625629411788, "brier": 0.4737080597217712, "mean_conf": 0.7311588013609494, "confident_error_rate": 0.0125, "coverage_at_0_9": 0.2125, "accuracy_at_0_9": 0.9411764705882353, "coverage_at_5pct_error": 0.1875, "coverage_at_1pct_error": 0.1875, "aurc": 0.16015706656418405, "error_rate_at_0_9": 0.058823529411764705, "confidence_bias": 0.08115880136094933, "top_bins": { "0.9": { "n": 17, "errors": 1, "error_rate": 0.058823529411764705 }, "0.95": { "n": 12, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.85, "confidence_cutoff": 0.7263951073590981 }, "0.8": { "coverage": 0.8, "accuracy": 0.75, "confidence_cutoff": 0.5548622064540625 } }, "score_mae": 0.45330552002269775, "ranked_probability_score": 0.07559174675562505 }, "banking77": { "n": 80, "nll": 0.3979839296889442, "acc": 0.875, "ece": 0.0922611325054158, "brier": 0.1822296579454367, "mean_conf": 0.8923134569904196, "confident_error_rate": 0.0375, "coverage_at_0_9": 0.725, "accuracy_at_0_9": 0.9482758620689655, "coverage_at_5pct_error": 0.825, "coverage_at_1pct_error": 0.675, "aurc": 0.02199412946893899, "error_rate_at_0_9": 0.05172413793103448, "confidence_bias": 0.01731345699041964, "top_bins": { "0.9": { "n": 58, "errors": 3, "error_rate": 0.05172413793103448 }, "0.95": { "n": 52, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 41, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9931638954155307 }, "0.8": { "coverage": 0.8, "accuracy": 0.953125, "confidence_cutoff": 0.845837989712747 } } }, "boolq": { "n": 80, "nll": 0.5220096518672049, "acc": 0.8125, "ece": 0.14833387100163806, "brier": 0.2951043064516692, "mean_conf": 0.9369053870550708, "confident_error_rate": 0.0875, "coverage_at_0_9": 0.8, "accuracy_at_0_9": 0.890625, "coverage_at_5pct_error": 0.35, "coverage_at_1pct_error": 0.2625, "aurc": 0.07004308805415525, "error_rate_at_0_9": 0.109375, "confidence_bias": 0.12440538705507076, "top_bins": { "0.9": { "n": 64, "errors": 7, "error_rate": 0.109375 }, "0.95": { "n": 57, "errors": 4, "error_rate": 0.07017543859649122 }, "0.99": { "n": 33, "errors": 3, "error_rate": 0.09090909090909091 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.925, "confidence_cutoff": 0.9818334678368117 }, "0.8": { "coverage": 0.8, "accuracy": 0.890625, "confidence_cutoff": 0.906916795952237 } } }, "composition_atom": { "n": 16, "nll": 0.52206201525939, "acc": 0.8125, "ece": 0.33102741792619084, "brier": 0.34727627266455574, "mean_conf": 0.678589841408767, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0625, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.5, "coverage_at_1pct_error": 0.5, "aurc": 0.1089238886113886, "error_rate_at_0_9": 0.0, "confidence_bias": -0.133910158591233, "top_bins": { "0.9": { "n": 1, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.7019797950393023 }, "0.8": { "coverage": 0.8125, "accuracy": 0.7692307692307693, "confidence_cutoff": 0.5925620021652579 } } }, "composition_conditional": { "n": 16, "nll": 0.7085893013145462, "acc": 0.5, "ece": 0.0392985537546805, "brier": 0.5153023405509731, "mean_conf": 0.5392985537546805, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.0, "coverage_at_1pct_error": 0.0, "aurc": 0.6738095238095239, "error_rate_at_0_9": null, "confidence_bias": 0.0392985537546805, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.25, "confidence_cutoff": 0.5389107305136105 }, "0.8": { "coverage": 0.875, "accuracy": 0.5714285714285714, "confidence_cutoff": 0.5247236050583006 } } }, "composition_conjunction": { "n": 16, "nll": 0.6954926808023254, "acc": 0.6875, "ece": 0.28356235633747884, "brier": 0.5005900249196042, "mean_conf": 0.5960339565951683, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.0, "coverage_at_1pct_error": 0.0, "aurc": 0.571116296897547, "error_rate_at_0_9": null, "confidence_bias": -0.09146604340483167, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.375, "confidence_cutoff": 0.5774681144198214 }, "0.8": { "coverage": 0.875, "accuracy": 0.6428571428571429, "confidence_cutoff": 0.5598945450814203 } } }, "composition_disjunction": { "n": 16, "nll": 0.6223727318080233, "acc": 0.5625, "ece": 0.22532601951230213, "brier": 0.4331353693805437, "mean_conf": 0.5895230696045874, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.3125, "coverage_at_1pct_error": 0.3125, "aurc": 0.22883184523809524, "error_rate_at_0_9": null, "confidence_bias": 0.027023069604587446, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5625, "accuracy": 0.6666666666666666, "confidence_cutoff": 0.5856707096099854 }, "0.8": { "coverage": 0.875, "accuracy": 0.6428571428571429, "confidence_cutoff": 0.5193371772766113 } } }, "composition_exception": { "n": 16, "nll": 0.6747223985633183, "acc": 0.6875, "ece": 0.16649871606180544, "brier": 0.4815999757579885, "mean_conf": 0.5210012839381946, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.0, "coverage_at_1pct_error": 0.0, "aurc": 0.3167455808080808, "error_rate_at_0_9": null, "confidence_bias": -0.16649871606180544, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5625, "accuracy": 0.7777777777777778, "confidence_cutoff": 0.518760085105896 }, "0.8": { "coverage": 0.875, "accuracy": 0.6428571428571429, "confidence_cutoff": 0.5035156757687234 } } }, "composition_negation": { "n": 16, "nll": 0.7403130749553563, "acc": 0.5, "ece": 0.08891041545292032, "brier": 0.5451303525712351, "mean_conf": 0.5889104154529203, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.0, "coverage_at_1pct_error": 0.0, "aurc": 0.6560387529137529, "error_rate_at_0_9": null, "confidence_bias": 0.08891041545292033, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.375, "confidence_cutoff": 0.5939869638673642 }, "0.8": { "coverage": 0.8125, "accuracy": 0.46153846153846156, "confidence_cutoff": 0.5687558393928909 } } }, "composition_nested_and": { "n": 16, "nll": 0.6701464137220343, "acc": 0.5, "ece": 0.1743572557807083, "brier": 0.4801171644212976, "mean_conf": 0.5974301225035049, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.25, "coverage_at_1pct_error": 0.25, "aurc": 0.37770632145632144, "error_rate_at_0_9": null, "confidence_bias": 0.09743012250350491, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.5, "confidence_cutoff": 0.5888296365737915 }, "0.8": { "coverage": 0.8125, "accuracy": 0.46153846153846156, "confidence_cutoff": 0.5511568948971299 } } }, "composition_nested_or": { "n": 16, "nll": 0.5812368508611689, "acc": 0.5625, "ece": 0.232265838198384, "brier": 0.39731880250019436, "mean_conf": 0.596988431257154, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.375, "coverage_at_1pct_error": 0.375, "aurc": 0.1931575369075369, "error_rate_at_0_9": null, "confidence_bias": 0.03448843125715395, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.875, "confidence_cutoff": 0.6048937439918518 }, "0.8": { "coverage": 0.8125, "accuracy": 0.6923076923076923, "confidence_cutoff": 0.5087750402458 } } }, "contrastive_age_eligibility": { "n": 24, "nll": 0.00030557725656388685, "acc": 1.0, "ece": 0.0003054692741715215, "brier": 4.3159530440543707e-07, "mean_conf": 0.9996945307258285, "confident_error_rate": 0.0, "coverage_at_0_9": 1.0, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 1.0, "coverage_at_1pct_error": 1.0, "aurc": 0.0, "error_rate_at_0_9": 0.0, "confidence_bias": -0.0003054692741715215, "top_bins": { "0.9": { "n": 24, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 24, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 24, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9998034213326141 }, "0.8": { "coverage": 0.8333333333333334, "accuracy": 1.0, "confidence_cutoff": 0.9996426136047368 } } }, "contrastive_quantity_limit": { "n": 24, "nll": 0.0949641863436191, "acc": 0.9583333333333334, "ece": 0.06630730027669939, "brier": 0.061578113407729874, "mean_conf": 0.9503996922670156, "confident_error_rate": 0.0, "coverage_at_0_9": 0.8333333333333334, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 1.0, "coverage_at_1pct_error": 0.9166666666666666, "aurc": 0.0035477053140096615, "error_rate_at_0_9": 0.0, "confidence_bias": -0.007933641066317776, "top_bins": { "0.9": { "n": 20, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 20, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 14, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9999154857901148 }, "0.8": { "coverage": 0.8333333333333334, "accuracy": 1.0, "confidence_cutoff": 0.9568661440200333 } } }, "contrastive_return_window": { "n": 24, "nll": 0.5242983013148222, "acc": 0.75, "ece": 0.13728604635195948, "brier": 0.3391539263874357, "mean_conf": 0.8872860463519596, "confident_error_rate": 0.08333333333333333, "coverage_at_0_9": 0.6666666666666666, "accuracy_at_0_9": 0.875, "coverage_at_5pct_error": 0.4583333333333333, "coverage_at_1pct_error": 0.4583333333333333, "aurc": 0.08703314878649715, "error_rate_at_0_9": 0.125, "confidence_bias": 0.13728604635195962, "top_bins": { "0.9": { "n": 16, "errors": 2, "error_rate": 0.125 }, "0.95": { "n": 11, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9166666666666666, "confidence_cutoff": 0.9489070242881575 }, "0.8": { "coverage": 0.8333333333333334, "accuracy": 0.8, "confidence_cutoff": 0.7711345531891699 } } }, "contrastive_spend_threshold": { "n": 24, "nll": 9.237405628065089e-06, "acc": 1.0, "ece": 9.237352093016682e-06, "brier": 2.021011459531347e-10, "mean_conf": 0.999990762647907, "confident_error_rate": 0.0, "coverage_at_0_9": 1.0, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 1.0, "coverage_at_1pct_error": 1.0, "aurc": 0.0, "error_rate_at_0_9": 0.0, "confidence_bias": -9.237352093016682e-06, "top_bins": { "0.9": { "n": 24, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 24, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 24, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9999922291820006 }, "0.8": { "coverage": 0.8333333333333334, "accuracy": 1.0, "confidence_cutoff": 0.9999883904450512 } } }, "dbpedia14": { "n": 80, "nll": 0.3415167120067256, "acc": 0.9625, "ece": 0.043700092546254746, "brier": 0.07457610595560957, "mean_conf": 0.9898401932810058, "confident_error_rate": 0.025, "coverage_at_0_9": 0.9625, "accuracy_at_0_9": 0.974025974025974, "coverage_at_5pct_error": 1.0, "coverage_at_1pct_error": 0.1375, "aurc": 0.02543183035490516, "error_rate_at_0_9": 0.025974025974025976, "confidence_bias": 0.027340193281005787, "top_bins": { "0.9": { "n": 77, "errors": 2, "error_rate": 0.025974025974025976 }, "0.95": { "n": 77, "errors": 2, "error_rate": 0.025974025974025976 }, "0.99": { "n": 77, "errors": 2, "error_rate": 0.025974025974025976 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.975, "confidence_cutoff": 0.9999976727949713 }, "0.8": { "coverage": 0.8, "accuracy": 0.984375, "confidence_cutoff": 0.999980934132584 } } }, "imdb": { "n": 80, "nll": 0.24985978507417891, "acc": 0.9375, "ece": 0.05457930309714365, "brier": 0.11189476167690375, "mean_conf": 0.977379170939226, "confident_error_rate": 0.05, "coverage_at_0_9": 0.95, "accuracy_at_0_9": 0.9473684210526315, "coverage_at_5pct_error": 0.9375, "coverage_at_1pct_error": 0.475, "aurc": 0.013302134857181047, "error_rate_at_0_9": 0.05263157894736842, "confidence_bias": 0.039879170939225994, "top_bins": { "0.9": { "n": 76, "errors": 4, "error_rate": 0.05263157894736842 }, "0.95": { "n": 74, "errors": 3, "error_rate": 0.04054054054054054 }, "0.99": { "n": 68, "errors": 1, "error_rate": 0.014705882352941176 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.975, "confidence_cutoff": 0.9992378023250749 }, "0.8": { "coverage": 0.8, "accuracy": 0.984375, "confidence_cutoff": 0.9947106490827637 } } }, "mnli": { "n": 80, "nll": 0.8964775304734388, "acc": 0.825, "ece": 0.16162789615757212, "brier": 0.33240517794745345, "mean_conf": 0.9328373683549026, "confident_error_rate": 0.1375, "coverage_at_0_9": 0.8, "accuracy_at_0_9": 0.828125, "coverage_at_5pct_error": 0.0625, "coverage_at_1pct_error": 0.0625, "aurc": 0.1279801828240891, "error_rate_at_0_9": 0.171875, "confidence_bias": 0.10783736835490265, "top_bins": { "0.9": { "n": 64, "errors": 11, "error_rate": 0.171875 }, "0.95": { "n": 62, "errors": 10, "error_rate": 0.16129032258064516 }, "0.99": { "n": 50, "errors": 7, "error_rate": 0.14 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.875, "confidence_cutoff": 0.9951284115930806 }, "0.8": { "coverage": 0.8, "accuracy": 0.828125, "confidence_cutoff": 0.9465963914191013 } } }, "sst5": { "n": 80, "nll": 1.0021265818153486, "acc": 0.6125, "ece": 0.11978396394038553, "brier": 0.5446736304135753, "mean_conf": 0.700510761326975, "confident_error_rate": 0.0125, "coverage_at_0_9": 0.075, "accuracy_at_0_9": 0.8333333333333334, "coverage_at_5pct_error": 0.0, "coverage_at_1pct_error": 0.0, "aurc": 0.3072321995862583, "error_rate_at_0_9": 0.16666666666666666, "confidence_bias": 0.08801076132697494, "top_bins": { "0.9": { "n": 6, "errors": 1, "error_rate": 0.16666666666666666 }, "0.95": { "n": 1, "errors": 1, "error_rate": 1.0 }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.675, "confidence_cutoff": 0.7306530300819783 }, "0.8": { "coverage": 0.8, "accuracy": 0.65625, "confidence_cutoff": 0.5438117284883917 } }, "score_mae": 0.5114106925506208, "ranked_probability_score": 0.08295468856515971 }, "trec": { "n": 80, "nll": 0.2925501809265985, "acc": 0.9375, "ece": 0.05267290937483785, "brier": 0.09750309073685226, "mean_conf": 0.9901729093748377, "confident_error_rate": 0.0375, "coverage_at_0_9": 0.975, "accuracy_at_0_9": 0.9615384615384616, "coverage_at_5pct_error": 0.975, "coverage_at_1pct_error": 0.65, "aurc": 0.010066847867815351, "error_rate_at_0_9": 0.038461538461538464, "confidence_bias": 0.052672909374837706, "top_bins": { "0.9": { "n": 78, "errors": 3, "error_rate": 0.038461538461538464 }, "0.95": { "n": 77, "errors": 2, "error_rate": 0.025974025974025976 }, "0.99": { "n": 74, "errors": 2, "error_rate": 0.02702702702702703 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9998811671798993 }, "0.8": { "coverage": 0.8, "accuracy": 0.96875, "confidence_cutoff": 0.9989895525699338 } } }, "yelp": { "n": 80, "nll": 0.724927007873615, "acc": 0.7125, "ece": 0.12184401403681792, "brier": 0.43845374961294875, "mean_conf": 0.7841603935118646, "confident_error_rate": 0.025, "coverage_at_0_9": 0.25, "accuracy_at_0_9": 0.9, "coverage_at_5pct_error": 0.1625, "coverage_at_1pct_error": 0.1625, "aurc": 0.18107778703006866, "error_rate_at_0_9": 0.1, "confidence_bias": 0.0716603935118646, "top_bins": { "0.9": { "n": 20, "errors": 2, "error_rate": 0.1 }, "0.95": { "n": 12, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 3, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.8, "confidence_cutoff": 0.8030074284504592 }, "0.8": { "coverage": 0.8, "accuracy": 0.734375, "confidence_cutoff": 0.6472819362414997 } }, "score_mae": 0.36393388420969347, "ranked_probability_score": 0.0574054980846211 }, "yelp_yn": { "n": 80, "nll": 0.1767395803682206, "acc": 0.925, "ece": 0.04299234681271804, "brier": 0.10751418779018915, "mean_conf": 0.9602990730093268, "confident_error_rate": 0.0125, "coverage_at_0_9": 0.8625, "accuracy_at_0_9": 0.9855072463768116, "coverage_at_5pct_error": 0.9125, "coverage_at_1pct_error": 0.6875, "aurc": 0.009561420798710252, "error_rate_at_0_9": 0.014492753623188406, "confidence_bias": 0.035299073009326776, "top_bins": { "0.9": { "n": 69, "errors": 1, "error_rate": 0.014492753623188406 }, "0.95": { "n": 64, "errors": 1, "error_rate": 0.015625 }, "0.99": { "n": 49, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9945756192344355 }, "0.8": { "coverage": 0.8, "accuracy": 0.984375, "confidence_cutoff": 0.9567033266284315 } } } }, "variants": { "clean": { "n": 1264, "nll": 0.4989297577819544, "acc": 0.8243670886075949, "ece": 0.05704976909178335, "brier": 0.25535485232026645, "mean_conf": 0.8785235750178558, "confident_error_rate": 0.040348101265822785, "coverage_at_0_9": 0.6566455696202531, "accuracy_at_0_9": 0.9385542168674699, "coverage_at_5pct_error": 0.6337025316455697, "coverage_at_1pct_error": 0.18908227848101267, "aurc": 0.057098658886616284, "error_rate_at_0_9": 0.06144578313253012, "confidence_bias": 0.05415648641026094, "top_bins": { "0.9": { "n": 830, "errors": 51, "error_rate": 0.06144578313253012 }, "0.95": { "n": 765, "errors": 32, "error_rate": 0.04183006535947712 }, "0.99": { "n": 616, "errors": 21, "error_rate": 0.03409090909090909 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9651898734177216, "confidence_cutoff": 0.9884021799746481 }, "0.8": { "coverage": 0.8006329113924051, "accuracy": 0.8913043478260869, "confidence_cutoff": 0.7007823799483796 } }, "score_mae": 0.44288336559433733, "ranked_probability_score": 0.07198397780180195 }, "none_absent": { "n": 60, "nll": 0.3973810280057748, "acc": 0.8666666666666667, "ece": 0.06408010262843164, "brier": 0.19542789202339048, "mean_conf": 0.9298361792245137, "confident_error_rate": 0.016666666666666666, "coverage_at_0_9": 0.7666666666666667, "accuracy_at_0_9": 0.9782608695652174, "coverage_at_5pct_error": 0.8, "coverage_at_1pct_error": 0.0, "aurc": 0.09434286048325893, "error_rate_at_0_9": 0.021739130434782608, "confidence_bias": 0.06316951255784697, "top_bins": { "0.9": { "n": 46, "errors": 1, "error_rate": 0.021739130434782608 }, "0.95": { "n": 40, "errors": 1, "error_rate": 0.025 }, "0.99": { "n": 23, "errors": 1, "error_rate": 0.043478260869565216 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9666666666666667, "confidence_cutoff": 0.9822965689718635 }, "0.8": { "coverage": 0.8, "accuracy": 0.9583333333333334, "confidence_cutoff": 0.8703627257812488 } } }, "none_present": { "n": 60, "nll": 0.8331048448864015, "acc": 0.85, "ece": 0.09958344599948121, "brier": 0.25491634205576813, "mean_conf": 0.8976962829199532, "confident_error_rate": 0.05, "coverage_at_0_9": 0.7, "accuracy_at_0_9": 0.9285714285714286, "coverage_at_5pct_error": 0.6666666666666666, "coverage_at_1pct_error": 0.1, "aurc": 0.0707109577846649, "error_rate_at_0_9": 0.07142857142857142, "confidence_bias": 0.0476962829199532, "top_bins": { "0.9": { "n": 42, "errors": 3, "error_rate": 0.07142857142857142 }, "0.95": { "n": 36, "errors": 2, "error_rate": 0.05555555555555555 }, "0.99": { "n": 25, "errors": 1, "error_rate": 0.04 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9333333333333333, "confidence_cutoff": 0.9711979045098904 }, "0.8": { "coverage": 0.8, "accuracy": 0.9166666666666666, "confidence_cutoff": 0.8156208519156563 } } }, "permuted": { "n": 84, "nll": 0.4638696687650698, "acc": 0.9166666666666666, "ece": 0.06166928073435357, "brier": 0.14839515349641946, "mean_conf": 0.9647966493434161, "confident_error_rate": 0.047619047619047616, "coverage_at_0_9": 0.9047619047619048, "accuracy_at_0_9": 0.9473684210526315, "coverage_at_5pct_error": 0.8214285714285714, "coverage_at_1pct_error": 0.05952380952380952, "aurc": 0.048011723747782474, "error_rate_at_0_9": 0.05263157894736842, "confidence_bias": 0.048129982676749505, "top_bins": { "0.9": { "n": 76, "errors": 4, "error_rate": 0.05263157894736842 }, "0.95": { "n": 72, "errors": 4, "error_rate": 0.05555555555555555 }, "0.99": { "n": 61, "errors": 3, "error_rate": 0.04918032786885246 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9761904761904762, "confidence_cutoff": 0.9974597361195856 }, "0.8": { "coverage": 0.8095238095238095, "accuracy": 0.9558823529411765, "confidence_cutoff": 0.9721480404501962 } } } }, "heldout_tasks": {}, "permutation": { "n": 60, "mean_max_delta": 0.018664570457196664, "flip_rate": 0.016666666666666666 }, "temperature": 1.5157165665103978, "calibrated_clean": { "n": 1264, "nll": 0.4416141187311443, "acc": 0.8243670886075949, "ece": 0.029334304090866264, "brier": 0.24694275137320498, "mean_conf": 0.8342071023922145, "confident_error_rate": 0.022943037974683545, "coverage_at_0_9": 0.567246835443038, "accuracy_at_0_9": 0.9595536959553695, "coverage_at_5pct_error": 0.622626582278481, "coverage_at_1pct_error": 0.21044303797468356, "aurc": 0.05692034604957053, "error_rate_at_0_9": 0.040446304044630406, "confidence_bias": 0.009840013784619628, "top_bins": { "0.9": { "n": 717, "errors": 29, "error_rate": 0.040446304044630406 }, "0.95": { "n": 615, "errors": 19, "error_rate": 0.030894308943089432 }, "0.99": { "n": 326, "errors": 6, "error_rate": 0.018404907975460124 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9651898734177216, "confidence_cutoff": 0.9445154395381901 }, "0.8": { "coverage": 0.8006329113924051, "accuracy": 0.8913043478260869, "confidence_cutoff": 0.606953327797692 } }, "score_mae": 0.4676383167623379, "ranked_probability_score": 0.07005212248660367 }, "metric_policy": { "version": 2, "selective_ties": "whole_confidence_groups", "coverage_at_error": "in-sample maximum over confidence thresholds; not a deployed error guarantee", "aurc": "right-step integral over whole confidence groups", "confident_error_rate": "high-confidence errors divided by all questions", "error_rate_at_0_9": "errors divided by questions accepted at p_max >= 0.9", "nll": "exact from logits when recorded; otherwise from floored probabilities", "nll_floor": 1e-09, "renormalize_returned_probabilities": true, "raw_sums_outside_1e_5": 0, "returned_zeros": 0 }, "coverage": { "requested_records": 1204, "requested_questions": 1468, "evaluated_records": 1204, "evaluated_questions": 1468, "rejected_records": 0, "truncated_records": 0 }, "latency_ms": { "median": 79.56905200080655, "p95": 175.74176955040457 }, "calibration": { "inference_temperature": 1.0, "additional_temperature": 1.5157165665103978, "logits_recorded": true }, "transfer": { "objective": -0.8434221557083209, "paired_flip": { "pairs": 64, "flip_rate": 0.40625, "both_correct_rate": 0.34375, "invariant_pairs": 24, "invariance_rate": 0.9583333333333334, "invariant_both_correct_rate": 0.5833333333333334 }, "unknowable": null, "clean": { "n": 656, "nll": 0.8332240113525762, "acc": 0.6021341463414634, "ece": 0.12798419967379476, "brier": 0.5023300109226222, "mean_conf": 0.7301183460152582, "confident_error_rate": 0.04573170731707317, "coverage_at_0_9": 0.3125, "accuracy_at_0_9": 0.8536585365853658, "coverage_at_5pct_error": 0.11890243902439024, "coverage_at_1pct_error": 0.09451219512195122, "aurc": 0.23059238555383313, "error_rate_at_0_9": 0.14634146341463414, "confidence_bias": 0.1279841996737947, "top_bins": { "0.9": { "n": 205, "errors": 30, "error_rate": 0.14634146341463414 }, "0.95": { "n": 151, "errors": 20, "error_rate": 0.13245033112582782 }, "0.99": { "n": 79, "errors": 4, "error_rate": 0.05063291139240506 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.725609756097561, "confidence_cutoff": 0.7209882521334812 }, "0.8": { "coverage": 0.8003048780487805, "accuracy": 0.6533333333333333, "confidence_cutoff": 0.5341488281217432 } }, "score_mae": 0.6641775577164577, "ranked_probability_score": 0.29565575671388794 }, "tasks": { "composition_held_and_or": { "n": 32, "nll": 0.6427335354097619, "acc": 0.65625, "ece": 0.07911568686379761, "brier": 0.45100153299560664, "mean_conf": 0.5864401989835899, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.15625, "coverage_at_1pct_error": 0.15625, "aurc": 0.24615223624507382, "error_rate_at_0_9": null, "confidence_bias": -0.0698098010164101, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.53125, "accuracy": 0.6470588235294118, "confidence_cutoff": 0.5770142502482459 }, "0.8": { "coverage": 0.8125, "accuracy": 0.6153846153846154, "confidence_cutoff": 0.5480732917785645 } } }, "composition_held_conditional": { "n": 32, "nll": 0.7002559076723711, "acc": 0.5, "ece": 0.040716346048888484, "brier": 0.5069657859898338, "mean_conf": 0.5407163460488885, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.0, "coverage_at_1pct_error": 0.0, "aurc": 0.5428605550039391, "error_rate_at_0_9": null, "confidence_bias": 0.04071634604888852, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.5, "confidence_cutoff": 0.5319679538002 }, "0.8": { "coverage": 0.8125, "accuracy": 0.46153846153846156, "confidence_cutoff": 0.516680121421814 } } }, "composition_held_or_not": { "n": 32, "nll": 0.6868598188008299, "acc": 0.5, "ece": 0.11919972571689182, "brier": 0.49389757171435983, "mean_conf": 0.5462136768533761, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.09375, "coverage_at_1pct_error": 0.09375, "aurc": 0.39094494868742846, "error_rate_at_0_9": null, "confidence_bias": 0.04621367685337607, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.6875, "confidence_cutoff": 0.5332706888479405 }, "0.8": { "coverage": 0.8125, "accuracy": 0.46153846153846156, "confidence_cutoff": 0.5239105664518307 } } }, "contrastive_authorization": { "n": 40, "nll": 0.1012225013321368, "acc": 0.975, "ece": 0.0636128189986208, "brier": 0.04847437535132595, "mean_conf": 0.9158894408229405, "confident_error_rate": 0.0, "coverage_at_0_9": 0.725, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 1.0, "coverage_at_1pct_error": 0.975, "aurc": 0.000625, "error_rate_at_0_9": 0.0, "confidence_bias": -0.05911055917705943, "top_bins": { "0.9": { "n": 29, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 23, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 20, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9909082970328174 }, "0.8": { "coverage": 0.8, "accuracy": 1.0, "confidence_cutoff": 0.8867185513663616 } } }, "contrastive_deadline": { "n": 40, "nll": 2.3533720238091194, "acc": 0.3, "ece": 0.6268044472682741, "brier": 1.1752809482099298, "mean_conf": 0.902081250744253, "confident_error_rate": 0.45, "coverage_at_0_9": 0.65, "accuracy_at_0_9": 0.3076923076923077, "coverage_at_5pct_error": 0.075, "coverage_at_1pct_error": 0.075, "aurc": 0.5450540740713118, "error_rate_at_0_9": 0.6923076923076923, "confidence_bias": 0.602081250744253, "top_bins": { "0.9": { "n": 26, "errors": 18, "error_rate": 0.6923076923076923 }, "0.95": { "n": 22, "errors": 14, "error_rate": 0.6363636363636364 }, "0.99": { "n": 9, "errors": 4, "error_rate": 0.4444444444444444 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.4, "confidence_cutoff": 0.9668620287297324 }, "0.8": { "coverage": 0.8, "accuracy": 0.3125, "confidence_cutoff": 0.8341740534671955 } }, "score_mae": 0.6641775577164577, "ranked_probability_score": 0.29565575671388794 }, "emotion": { "n": 80, "nll": 1.1638223380813226, "acc": 0.525, "ece": 0.11298484196696625, "brier": 0.5903870224663738, "mean_conf": 0.6261382515881248, "confident_error_rate": 0.0125, "coverage_at_0_9": 0.1125, "accuracy_at_0_9": 0.8888888888888888, "coverage_at_5pct_error": 0.0625, "coverage_at_1pct_error": 0.0625, "aurc": 0.2782174494810715, "error_rate_at_0_9": 0.1111111111111111, "confidence_bias": 0.10113825158812473, "top_bins": { "0.9": { "n": 9, "errors": 1, "error_rate": 0.1111111111111111 }, "0.95": { "n": 5, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 1, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.725, "confidence_cutoff": 0.6017182592654616 }, "0.8": { "coverage": 0.8, "accuracy": 0.609375, "confidence_cutoff": 0.4732868788719531 } } }, "mmlu": { "n": 80, "nll": 1.2132248805897294, "acc": 0.425, "ece": 0.17901472294834253, "brier": 0.6687983463380192, "mean_conf": 0.5761396137039483, "confident_error_rate": 0.0, "coverage_at_0_9": 0.125, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.15, "coverage_at_1pct_error": 0.15, "aurc": 0.38430439973401825, "error_rate_at_0_9": 0.0, "confidence_bias": 0.1511396137039483, "top_bins": { "0.9": { "n": 10, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 8, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 5, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.6, "confidence_cutoff": 0.5105295219311827 }, "0.8": { "coverage": 0.8, "accuracy": 0.4375, "confidence_cutoff": 0.385735625241166 } } }, "paws": { "n": 80, "nll": 1.00533805602315, "acc": 0.3, "ece": 0.3806088203975223, "brier": 0.7785508948914812, "mean_conf": 0.6776616149008904, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0125, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.0375, "coverage_at_1pct_error": 0.0375, "aurc": 0.7547032981712617, "error_rate_at_0_9": 0.0, "confidence_bias": 0.3776616149008904, "top_bins": { "0.9": { "n": 1, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.15, "confidence_cutoff": 0.6846514736219801 }, "0.8": { "coverage": 0.8, "accuracy": 0.25, "confidence_cutoff": 0.580651581287384 } } }, "qnli": { "n": 80, "nll": 0.5594003098802613, "acc": 0.7625, "ece": 0.10667716921885279, "brier": 0.3553818939712926, "mean_conf": 0.818959518366006, "confident_error_rate": 0.0625, "coverage_at_0_9": 0.4375, "accuracy_at_0_9": 0.8571428571428571, "coverage_at_5pct_error": 0.0625, "coverage_at_1pct_error": 0.0625, "aurc": 0.14924964619005282, "error_rate_at_0_9": 0.14285714285714285, "confidence_bias": 0.056459518366006045, "top_bins": { "0.9": { "n": 35, "errors": 5, "error_rate": 0.14285714285714285 }, "0.95": { "n": 21, "errors": 2, "error_rate": 0.09523809523809523 }, "0.99": { "n": 4, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.825, "confidence_cutoff": 0.8597152316740018 }, "0.8": { "coverage": 0.8, "accuracy": 0.78125, "confidence_cutoff": 0.675449788570404 } } }, "sciq": { "n": 80, "nll": 0.18838356801169662, "acc": 0.925, "ece": 0.06864685652623173, "brier": 0.10770728627230633, "mean_conf": 0.895637017961667, "confident_error_rate": 0.0, "coverage_at_0_9": 0.7625, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.95, "coverage_at_1pct_error": 0.7875, "aurc": 0.007778982639311541, "error_rate_at_0_9": 0.0, "confidence_bias": -0.029362982038333096, "top_bins": { "0.9": { "n": 61, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 54, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 40, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9910152874950432 }, "0.8": { "coverage": 0.8, "accuracy": 0.984375, "confidence_cutoff": 0.8018387376851126 } } }, "tweet_offensive": { "n": 80, "nll": 0.6630307731811524, "acc": 0.7, "ece": 0.11410098626654272, "brier": 0.4256570275654813, "mean_conf": 0.8141009862665427, "confident_error_rate": 0.075, "coverage_at_0_9": 0.425, "accuracy_at_0_9": 0.8235294117647058, "coverage_at_5pct_error": 0.0875, "coverage_at_1pct_error": 0.0875, "aurc": 0.20144605524801446, "error_rate_at_0_9": 0.17647058823529413, "confidence_bias": 0.11410098626654275, "top_bins": { "0.9": { "n": 34, "errors": 6, "error_rate": 0.17647058823529413 }, "0.95": { "n": 18, "errors": 4, "error_rate": 0.2222222222222222 }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.825, "confidence_cutoff": 0.8868192435698719 }, "0.8": { "coverage": 0.8, "accuracy": 0.75, "confidence_cutoff": 0.6436328483710542 } } } }, "variants": { "clean": { "n": 656, "nll": 0.8332240113525762, "acc": 0.6021341463414634, "ece": 0.12798419967379476, "brier": 0.5023300109226222, "mean_conf": 0.7301183460152582, "confident_error_rate": 0.04573170731707317, "coverage_at_0_9": 0.3125, "accuracy_at_0_9": 0.8536585365853658, "coverage_at_5pct_error": 0.11890243902439024, "coverage_at_1pct_error": 0.09451219512195122, "aurc": 0.23059238555383313, "error_rate_at_0_9": 0.14634146341463414, "confidence_bias": 0.1279841996737947, "top_bins": { "0.9": { "n": 205, "errors": 30, "error_rate": 0.14634146341463414 }, "0.95": { "n": 151, "errors": 20, "error_rate": 0.13245033112582782 }, "0.99": { "n": 79, "errors": 4, "error_rate": 0.05063291139240506 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.725609756097561, "confidence_cutoff": 0.7209882521334812 }, "0.8": { "coverage": 0.8003048780487805, "accuracy": 0.6533333333333333, "confidence_cutoff": 0.5341488281217432 } }, "score_mae": 0.6641775577164577, "ranked_probability_score": 0.29565575671388794 }, "none_absent": { "n": 36, "nll": 0.6685620869807386, "acc": 0.7777777777777778, "ece": 0.18033089059255358, "brier": 0.37205937297465375, "mean_conf": 0.7657635384076253, "confident_error_rate": 0.08333333333333333, "coverage_at_0_9": 0.4166666666666667, "accuracy_at_0_9": 0.8, "coverage_at_5pct_error": 0.08333333333333333, "coverage_at_1pct_error": 0.08333333333333333, "aurc": 0.17909033623816945, "error_rate_at_0_9": 0.2, "confidence_bias": -0.012014239370152446, "top_bins": { "0.9": { "n": 15, "errors": 3, "error_rate": 0.2 }, "0.95": { "n": 6, "errors": 1, "error_rate": 0.16666666666666666 }, "0.99": { "n": 2, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.7777777777777778, "confidence_cutoff": 0.8402871256038218 }, "0.8": { "coverage": 0.8055555555555556, "accuracy": 0.7931034482758621, "confidence_cutoff": 0.5899206246083812 } } }, "none_present": { "n": 36, "nll": 1.2061311893085476, "acc": 0.5833333333333334, "ece": 0.24336477102423715, "brier": 0.5794150123852385, "mean_conf": 0.7670494004663645, "confident_error_rate": 0.0, "coverage_at_0_9": 0.3333333333333333, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.4166666666666667, "coverage_at_1pct_error": 0.4166666666666667, "aurc": 0.1664477209679199, "error_rate_at_0_9": 0.0, "confidence_bias": 0.18371606713303112, "top_bins": { "0.9": { "n": 12, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 9, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 2, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.8888888888888888, "confidence_cutoff": 0.8493899555479588 }, "0.8": { "coverage": 0.8055555555555556, "accuracy": 0.6551724137931034, "confidence_cutoff": 0.5037377697189107 } } }, "permuted": { "n": 36, "nll": 0.710652540300337, "acc": 0.75, "ece": 0.16435148952453013, "brier": 0.37681908989783053, "mean_conf": 0.7630775644796315, "confident_error_rate": 0.0, "coverage_at_0_9": 0.4166666666666667, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.4444444444444444, "coverage_at_1pct_error": 0.4444444444444444, "aurc": 0.09547117668697218, "error_rate_at_0_9": 0.0, "confidence_bias": 0.013077564479631487, "top_bins": { "0.9": { "n": 15, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 15, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 8, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.9444444444444444, "confidence_cutoff": 0.8530542285676821 }, "0.8": { "coverage": 0.8055555555555556, "accuracy": 0.7586206896551724, "confidence_cutoff": 0.4821188982489635 } } } }, "heldout_tasks": { "composition_held_and_or": { "n": 32, "nll": 0.6427335354097619, "acc": 0.65625, "ece": 0.07911568686379761, "brier": 0.45100153299560664, "mean_conf": 0.5864401989835899, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.15625, "coverage_at_1pct_error": 0.15625, "aurc": 0.24615223624507382, "error_rate_at_0_9": null, "confidence_bias": -0.0698098010164101, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.53125, "accuracy": 0.6470588235294118, "confidence_cutoff": 0.5770142502482459 }, "0.8": { "coverage": 0.8125, "accuracy": 0.6153846153846154, "confidence_cutoff": 0.5480732917785645 } } }, "composition_held_conditional": { "n": 32, "nll": 0.7002559076723711, "acc": 0.5, "ece": 0.040716346048888484, "brier": 0.5069657859898338, "mean_conf": 0.5407163460488885, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.0, "coverage_at_1pct_error": 0.0, "aurc": 0.5428605550039391, "error_rate_at_0_9": null, "confidence_bias": 0.04071634604888852, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.5, "confidence_cutoff": 0.5319679538002 }, "0.8": { "coverage": 0.8125, "accuracy": 0.46153846153846156, "confidence_cutoff": 0.516680121421814 } } }, "composition_held_or_not": { "n": 32, "nll": 0.6868598188008299, "acc": 0.5, "ece": 0.11919972571689182, "brier": 0.49389757171435983, "mean_conf": 0.5462136768533761, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0, "accuracy_at_0_9": null, "coverage_at_5pct_error": 0.09375, "coverage_at_1pct_error": 0.09375, "aurc": 0.39094494868742846, "error_rate_at_0_9": null, "confidence_bias": 0.04621367685337607, "top_bins": { "0.9": { "n": 0, "errors": 0, "error_rate": null }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.6875, "confidence_cutoff": 0.5332706888479405 }, "0.8": { "coverage": 0.8125, "accuracy": 0.46153846153846156, "confidence_cutoff": 0.5239105664518307 } } }, "contrastive_authorization": { "n": 40, "nll": 0.1012225013321368, "acc": 0.975, "ece": 0.0636128189986208, "brier": 0.04847437535132595, "mean_conf": 0.9158894408229405, "confident_error_rate": 0.0, "coverage_at_0_9": 0.725, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 1.0, "coverage_at_1pct_error": 0.975, "aurc": 0.000625, "error_rate_at_0_9": 0.0, "confidence_bias": -0.05911055917705943, "top_bins": { "0.9": { "n": 29, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 23, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 20, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9909082970328174 }, "0.8": { "coverage": 0.8, "accuracy": 1.0, "confidence_cutoff": 0.8867185513663616 } } }, "contrastive_deadline": { "n": 40, "nll": 2.3533720238091194, "acc": 0.3, "ece": 0.6268044472682741, "brier": 1.1752809482099298, "mean_conf": 0.902081250744253, "confident_error_rate": 0.45, "coverage_at_0_9": 0.65, "accuracy_at_0_9": 0.3076923076923077, "coverage_at_5pct_error": 0.075, "coverage_at_1pct_error": 0.075, "aurc": 0.5450540740713118, "error_rate_at_0_9": 0.6923076923076923, "confidence_bias": 0.602081250744253, "top_bins": { "0.9": { "n": 26, "errors": 18, "error_rate": 0.6923076923076923 }, "0.95": { "n": 22, "errors": 14, "error_rate": 0.6363636363636364 }, "0.99": { "n": 9, "errors": 4, "error_rate": 0.4444444444444444 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.4, "confidence_cutoff": 0.9668620287297324 }, "0.8": { "coverage": 0.8, "accuracy": 0.3125, "confidence_cutoff": 0.8341740534671955 } }, "score_mae": 0.6641775577164577, "ranked_probability_score": 0.29565575671388794 }, "emotion": { "n": 80, "nll": 1.1638223380813226, "acc": 0.525, "ece": 0.11298484196696625, "brier": 0.5903870224663738, "mean_conf": 0.6261382515881248, "confident_error_rate": 0.0125, "coverage_at_0_9": 0.1125, "accuracy_at_0_9": 0.8888888888888888, "coverage_at_5pct_error": 0.0625, "coverage_at_1pct_error": 0.0625, "aurc": 0.2782174494810715, "error_rate_at_0_9": 0.1111111111111111, "confidence_bias": 0.10113825158812473, "top_bins": { "0.9": { "n": 9, "errors": 1, "error_rate": 0.1111111111111111 }, "0.95": { "n": 5, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 1, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.725, "confidence_cutoff": 0.6017182592654616 }, "0.8": { "coverage": 0.8, "accuracy": 0.609375, "confidence_cutoff": 0.4732868788719531 } } }, "mmlu": { "n": 80, "nll": 1.2132248805897294, "acc": 0.425, "ece": 0.17901472294834253, "brier": 0.6687983463380192, "mean_conf": 0.5761396137039483, "confident_error_rate": 0.0, "coverage_at_0_9": 0.125, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.15, "coverage_at_1pct_error": 0.15, "aurc": 0.38430439973401825, "error_rate_at_0_9": 0.0, "confidence_bias": 0.1511396137039483, "top_bins": { "0.9": { "n": 10, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 8, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 5, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.6, "confidence_cutoff": 0.5105295219311827 }, "0.8": { "coverage": 0.8, "accuracy": 0.4375, "confidence_cutoff": 0.385735625241166 } } }, "paws": { "n": 80, "nll": 1.00533805602315, "acc": 0.3, "ece": 0.3806088203975223, "brier": 0.7785508948914812, "mean_conf": 0.6776616149008904, "confident_error_rate": 0.0, "coverage_at_0_9": 0.0125, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.0375, "coverage_at_1pct_error": 0.0375, "aurc": 0.7547032981712617, "error_rate_at_0_9": 0.0, "confidence_bias": 0.3776616149008904, "top_bins": { "0.9": { "n": 1, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 0, "errors": 0, "error_rate": null }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.15, "confidence_cutoff": 0.6846514736219801 }, "0.8": { "coverage": 0.8, "accuracy": 0.25, "confidence_cutoff": 0.580651581287384 } } }, "qnli": { "n": 80, "nll": 0.5594003098802613, "acc": 0.7625, "ece": 0.10667716921885279, "brier": 0.3553818939712926, "mean_conf": 0.818959518366006, "confident_error_rate": 0.0625, "coverage_at_0_9": 0.4375, "accuracy_at_0_9": 0.8571428571428571, "coverage_at_5pct_error": 0.0625, "coverage_at_1pct_error": 0.0625, "aurc": 0.14924964619005282, "error_rate_at_0_9": 0.14285714285714285, "confidence_bias": 0.056459518366006045, "top_bins": { "0.9": { "n": 35, "errors": 5, "error_rate": 0.14285714285714285 }, "0.95": { "n": 21, "errors": 2, "error_rate": 0.09523809523809523 }, "0.99": { "n": 4, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.825, "confidence_cutoff": 0.8597152316740018 }, "0.8": { "coverage": 0.8, "accuracy": 0.78125, "confidence_cutoff": 0.675449788570404 } } }, "sciq": { "n": 80, "nll": 0.18838356801169662, "acc": 0.925, "ece": 0.06864685652623173, "brier": 0.10770728627230633, "mean_conf": 0.895637017961667, "confident_error_rate": 0.0, "coverage_at_0_9": 0.7625, "accuracy_at_0_9": 1.0, "coverage_at_5pct_error": 0.95, "coverage_at_1pct_error": 0.7875, "aurc": 0.007778982639311541, "error_rate_at_0_9": 0.0, "confidence_bias": -0.029362982038333096, "top_bins": { "0.9": { "n": 61, "errors": 0, "error_rate": 0.0 }, "0.95": { "n": 54, "errors": 0, "error_rate": 0.0 }, "0.99": { "n": 40, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 1.0, "confidence_cutoff": 0.9910152874950432 }, "0.8": { "coverage": 0.8, "accuracy": 0.984375, "confidence_cutoff": 0.8018387376851126 } } }, "tweet_offensive": { "n": 80, "nll": 0.6630307731811524, "acc": 0.7, "ece": 0.11410098626654272, "brier": 0.4256570275654813, "mean_conf": 0.8141009862665427, "confident_error_rate": 0.075, "coverage_at_0_9": 0.425, "accuracy_at_0_9": 0.8235294117647058, "coverage_at_5pct_error": 0.0875, "coverage_at_1pct_error": 0.0875, "aurc": 0.20144605524801446, "error_rate_at_0_9": 0.17647058823529413, "confidence_bias": 0.11410098626654275, "top_bins": { "0.9": { "n": 34, "errors": 6, "error_rate": 0.17647058823529413 }, "0.95": { "n": 18, "errors": 4, "error_rate": 0.2222222222222222 }, "0.99": { "n": 0, "errors": 0, "error_rate": null } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.825, "confidence_cutoff": 0.8868192435698719 }, "0.8": { "coverage": 0.8, "accuracy": 0.75, "confidence_cutoff": 0.6436328483710542 } } } }, "permutation": { "n": 36, "mean_max_delta": 0.07456167804296458, "flip_rate": 0.1111111111111111 }, "temperature": 1.5157165665103978, "calibrated_clean": { "n": 656, "nll": 0.7627521342590968, "acc": 0.6021341463414634, "ece": 0.07572062366206761, "brier": 0.4747254916137418, "mean_conf": 0.669677827611837, "confident_error_rate": 0.021341463414634148, "coverage_at_0_9": 0.18445121951219512, "accuracy_at_0_9": 0.8842975206611571, "coverage_at_5pct_error": 0.12652439024390244, "coverage_at_1pct_error": 0.09298780487804878, "aurc": 0.23002342930643124, "error_rate_at_0_9": 0.11570247933884298, "confidence_bias": 0.06754368127037358, "top_bins": { "0.9": { "n": 121, "errors": 14, "error_rate": 0.11570247933884298 }, "0.95": { "n": 75, "errors": 4, "error_rate": 0.05333333333333334 }, "0.99": { "n": 15, "errors": 0, "error_rate": 0.0 } }, "selective": { "0.5": { "coverage": 0.5, "accuracy": 0.7164634146341463, "confidence_cutoff": 0.641505976392506 }, "0.8": { "coverage": 0.8003048780487805, "accuracy": 0.6476190476190476, "confidence_cutoff": 0.5125548604205254 } }, "score_mae": 0.6487109061534533, "ranked_probability_score": 0.2651482098948402 }, "metric_policy": { "version": 2, "selective_ties": "whole_confidence_groups", "coverage_at_error": "in-sample maximum over confidence thresholds; not a deployed error guarantee", "aurc": "right-step integral over whole confidence groups", "confident_error_rate": "high-confidence errors divided by all questions", "error_rate_at_0_9": "errors divided by questions accepted at p_max >= 0.9", "nll": "exact from logits when recorded; otherwise from floored probabilities", "nll_floor": 1e-09, "renormalize_returned_probabilities": true, "raw_sums_outside_1e_5": 0, "returned_zeros": 0 }, "coverage": { "requested_records": 764, "requested_questions": 764, "evaluated_records": 764, "evaluated_questions": 764, "rejected_records": 0, "truncated_records": 0 }, "latency_ms": { "median": 79.30278050116613, "p95": 83.82728769975074 }, "calibration": { "inference_temperature": 1.0, "additional_temperature": 1.5157165665103978, "logits_recorded": true }, "suite_sha256": "31677c2256b406222e7d94ffdc0a02a70ce05746b9efe307876024c4e77291d1" }, "provenance": { "config": { "epochs": 2, "seed": 0, "lr": 0.0001, "lora": 16, "accum": 2, "batch": 4, "perm_kl": 0.0, "perm_frac": 0.3, "ord_w": 0.0, "p_none": 0.1, "p_none_distract": 0.12, "p_distract": 0.15, "p_none_pair": 0.25, "synthetic_repeat": 1, "public_frac": 1.0, "head_lr": 0.0, "weight_decay": 0.01, "anchor_w": 0.0, "label_smoothing": 0.0, "brier_w": 0.0, "focal_gamma": 0.0, "base": "google/gemma-4-E2B", "base_revision": "d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f", "dtype": "bf16", "weights_dtype": "bf16", "checkpointing": 1 }, "config_sha256": "cb5576f6ebf3fffe10370041387b4b6810c03c2b112b925168c16ded6729c34c", "suite_sha256": "a8f50e481b7d90b97da049e0ff6a01cee2f1ed204aed61a8265af0edbb5514d2", "source_hashes": { "kev/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "kev/anchors.py": "6963eafdb276db5a6c94d939eae675c761555553b8d0f963ffd803448beaabbb", "kev/api.py": "7bffacfb762c626b8bc2f670f350295af5ccb0883dbe90239c7d2f8e5ef56582", "kev/autoresearch.py": "f7d7fabb9c0df7065bee3fec4aa4bee7028c5d1565847aefe2d74e3d7d40ed8a", "kev/benchmark.py": "9c442efd8b6b852c574a646192c61ee91afa1306e3b2978bb6391ae1b91a3fa0", "kev/calibrate.py": "c1eff4744fabd349e8abca86777a7aa0cbea44904c8223d3b61fca3d43741519", "kev/checkpoint.py": "9f1dbc5d3d7d3050aca4ee34e0578c9ad76b7a2a7467dfc7a4ad65e3c471db59", "kev/compare.py": "3606dbf98bb305d66420158498cb837304d2edb8dab01cdaf2ec42964bef3435", "kev/composition.py": "f335ed17e18e0a544893db5e22b9a059e6ce1b2e14dbb863ac7d7bbf8f3e0536", "kev/contrastive.py": "cbb979aa5d40265ad0e64695f94b281d91751fa405811ddfa8212fede111edcf", "kev/data.py": "ef1be396e7218ba5c94f0f67de2ceff23a39656b57199cb3aa3192cde53e8320", "kev/device.py": "d1677fd98ec0979c7284546306e34e0d09298ca4042fc39eb2b32a74f4e975c4", "kev/evaluate.py": "999264f2837dcfbf2ec93601aa4745e701698ab850a43ad89324672a6604965c", "kev/experiment.py": "d12ab7f0765de0f29364f3411b99f06bb80b1cbf120ca9ed2b03316ec766e828", "kev/jev.py": "acd4cc3f1844e438cc83a8d409c15ef78a5ef64d39ea10f583d75a6c7646b243", "kev/metrics.py": "dba8d90550999edcd642581fb4da0367bd4b39d08471815184fa732d11e137b9", "kev/mlx_model.py": "f582428796faf6bf962b772a69a6227ac3259ccbd3215c0c8b55ce28dbbd0909", "kev/model.py": "c2129026ab56395f78c7d3071410541d5423dd7d97f7371518570572f206c376", "kev/plot.py": "0874bcff2885d8155a1cceade0de8a275a7163d3c3fd6aa0c294e8ccf5ff7e02", "kev/predictors.py": "b2a66dd9f0f0a8026bc94603383bacfdd622e75d1c935173997175606626a9b3", "kev/publish.py": "8abc9bcf4a01365697b05cf3c5ad0013462bd92d005f5616954e40ab1100c7aa", "kev/serve.py": "c210b1f4598e64b49c399f73603ed372f4314c165ceda50e1340c0466ec1c735", "kev/study_v3.py": "7891150aa623e4479c7c789d4dc4662f186132b37b7bac1e1f072b9e08ee13e3", "kev/suite.py": "d18d5f4f02d079344d61c2a31882c524c90012ebfdb660bf9160e145f37e3d42", "kev/train.py": "e103e99d3a68ab99eb58f135320fbc46b18514fcc2b195d1c6879633836e99b9", "kev/transfer_v9.py": "0902409742151250a28af1fd8f42258b70fcc161f3deed3ee35fe3f473f3c763", "modal_app.py": "7a13e9d29b3fc68f8236a0c1e4d67079b40a2a328a296f7ced1022817dff82aa", "pyproject.toml": "7c17fbe9efcadda3eb488b59adf1b6db5938cd33af4d3bd389ce4359b73fcc7d", "uv.lock": "a9922dbb89acdef78299fd2b4a8c3f7f0fa1b2bc08b55595b6926fa785a9c466" }, "git_commit": "ccd6e8b646068a6689792d76e3cae3698e9a015a", "platform": "Linux-6.12.103-127.188.amzn2023.x86_64-x86_64-with-glibc2.36", "torch": "2.8.0+cu128", "device": "cuda", "gpu": "NVIDIA L4", "legacy_checkpoint": false, "measured_checkpoint": { "requested": "/kev/runs/g4-e2b-e2-l4/00-trial-0/checkpoint", "resolved": "/kev/runs/g4-e2b-e2-l4/00-trial-0/checkpoint", "head_sha256": "dd2d8c5b1ac8bee3f413cd89802d93826674baf181dd4649004868b358605785", "adapter_sha256": "e23dbe51453b6e926da0db70a38d680bc64644caff872cf0defcbbd8b1acbe46", "inference_temperature": 1.0 } }, "mechanism_checks": { "n": 8, "packed_max_delta": 0.0, "sibling_max_delta": 0.0069029927253723145, "tolerance": 0.001, "passed": false }, "gates": { "passed": false, "checks": { "complete_coverage": true, "isolation_and_packing": false, "transfer_complete": true, "heldout_pairs_at_least_70pct": false, "transfer_confident_errors_below_10pct": true }, "policy": "Research screening only: 0.001 isolation; 5pp task regression; 70% heldout pair correctness; <=10% confident errors; no automatic release." }, "calibration_fit": { "temperature": 1.5157165665103978, "aggregation": "micro", "objective": "raw-logit NLL", "split": "calibration", "rows_sha256": "d51919a6df076619eb78186da47cebbf6b669444a4c356a5ff4a244184fa6433", "suite_sha256": "a8f50e481b7d90b97da049e0ff6a01cee2f1ed204aed61a8265af0edbb5514d2", "n": 1148 }, "wall_seconds": 10031.380025413, "promotable": false, "test_evaluated": false, "training_resources": { "wall_seconds": 9709.689131975174, "records_seen": 30212, "requested_records": 25012, "truncated_records": 0, "rejected_records": 0, "optimizer_steps": 3128, "forward_tokens": 6112066, "peak_device_bytes": 14624471040, "device": "cuda", "dtype": "bf16", "batch": 4, "peak_rss_bytes": 11347705856 } }