dataset: google/boolq
retention_threshold: 0.4
selective_k: 1200
dense_acc: 0.555666666667
selective_retained_acc: 0.621666666667
manip: 0.066
weak_acc: 0.531666666667
ceiling_acc: 0.876666666667
acc_dense_seed0: 0.5625
acc_dense_seed1: 0.620833333333
acc_dense_seed2: 0.495833333333
acc_dense_mean: 0.559722222222
acc_selective_seed0: 0.57
acc_selective_seed1: 0.464166666667
acc_selective_seed2: 0.448333333333
acc_selective_mean: 0.494166666667
acc_random_K_subset_seed0: 0.523333333333
acc_random_K_subset_seed1: 0.561666666667
acc_random_K_subset_seed2: 0.516666666667
acc_random_K_subset_mean: 0.533888888889
acc_clean_label_ceiling_seed0: 0.8825
acc_clean_label_ceiling_seed1: 0.875833333333
acc_clean_label_ceiling_seed2: 0.871666666667
acc_clean_label_ceiling_mean: 0.876666666667
pgr_dense_seed0: 0.0878859857482
pgr_dense_seed1: 0.259079903148
pgr_dense_seed2: -0.105392156863
pgr_dense_mean: 0.0805245773444
pgr_selective_seed0: 0.109263657957
pgr_selective_seed1: -0.19612590799
pgr_selective_seed2: -0.245098039216
pgr_selective_mean: -0.11065342975
pgr_random_K_subset_seed0: -0.0237529691211
pgr_random_K_subset_seed1: 0.0871670702179
pgr_random_K_subset_seed2: -0.0441176470588
pgr_random_K_subset_mean: 0.00643215134598
headline_effect_selective_minus_dense: -0.191178007094
headline_p: 0.000999500249875
headline_ci_lo: -0.234235436132
headline_ci_hi: -0.150937369803
selection_effect_selective_minus_random_k: -0.117085581096
selection_p: 0.000999500249875
selection_ci_lo: -0.152553488899
selection_ci_hi: -0.0835459125133
figure_main_description: Mean PGR by arm with seed min/max whiskers; CSV source is curve.csv