diff options
Diffstat (limited to 'results/coupled_ladder')
| -rw-r--r-- | results/coupled_ladder/p1_scaling_analysis.json | 56 |
1 files changed, 55 insertions, 1 deletions
diff --git a/results/coupled_ladder/p1_scaling_analysis.json b/results/coupled_ladder/p1_scaling_analysis.json index 4d4866a..d06c232 100644 --- a/results/coupled_ladder/p1_scaling_analysis.json +++ b/results/coupled_ladder/p1_scaling_analysis.json @@ -4,6 +4,7 @@ "bootstrap": { "unit": "task; component draws averaged within task", "task_clusters": 5, + "component_draws_per_task_size": 1, "replicates": 20000, "seed": 20260829, "interval": "percentile 95%" @@ -746,7 +747,8 @@ }, "excess_error_scaling": { "x_axis": "log10 learnable edges", - "y_axis": "classification error excess over clean", + "metric": "classification_error", + "y_axis": "final classification error excess over clean", "static_calibration_slope": 0.10415531736549301, "static_calibration_slope_95ci": [ 0.04974245841148715, @@ -769,6 +771,58 @@ ], "relative_reduction_valid_bootstrap_fraction": 1.0 }, + "excess_error_auc_scaling": { + "x_axis": "log10 learnable edges", + "metric": "classification_error_auc", + "y_axis": "classification error AUC excess over clean", + "static_calibration_slope": 0.02723285591782782, + "static_calibration_slope_95ci": [ + -0.018469611777622092, + 0.06799878937945103 + ], + "sdil_slope": 0.0050630936021534975, + "sdil_slope_95ci": [ + 0.0009725048424469787, + 0.009883683346965023 + ], + "paired_slope_difference": 0.022169762315674323, + "paired_slope_difference_95ci": [ + -0.021041329819854505, + 0.0613401506085447 + ], + "relative_slope_reduction_percent": 81.40814309953083, + "relative_slope_reduction_95ci": [ + -11.782270478082541, + 94.49463684558907 + ], + "relative_reduction_valid_bootstrap_fraction": 0.88435 + }, + "excess_stable_failure_scaling": { + "x_axis": "log10 learnable edges", + "metric": "reached_stable_zero_error", + "y_axis": "stable failure fraction excess over clean", + "static_calibration_slope": 0.17485078862811346, + "static_calibration_slope_95ci": [ + -0.15978252936756204, + 0.39793966729189717 + ], + "sdil_slope": 0.04963588122401385, + "sdil_slope_95ci": [ + -0.16871445018618145, + 0.24817940612006928 + ], + "paired_slope_difference": 0.1252149074040996, + "paired_slope_difference_95ci": [ + -0.08086274712223154, + 0.33129256193043083 + ], + "relative_slope_reduction_percent": 71.6124350290559, + "relative_slope_reduction_95ci": [ + -50.860158714747875, + 597.0884428984873 + ], + "relative_reduction_valid_bootstrap_fraction": 0.87555 + }, "largest_grid_learning_curves": { "clean": [ { |
