pnw-subtask-benchmark / results /calibration.json
tomas-duarte's picture
P&W subtask segmentation benchmark
63f0b54
Raw History Blame Contribute Delete
1.44 kB
{
"matched_attempts": 3843,
"predicted_attempts": 4156,
"ece_10bin": 0.0425,
"brier": 0.0341,
"share_below_10_or_above_90": 0.7299,
"success_rate": 0.5857,
"mean_p": 0.6076,
"reliability": [
{
"bin": [
0.0,
0.1
],
"n": 899,
"mean_p": 0.065,
"success_rate": 0.0056
},
{
"bin": [
0.1,
0.2
],
"n": 442,
"mean_p": 0.1357,
"success_rate": 0.0271
},
{
"bin": [
0.2,
0.3
],
"n": 120,
"mean_p": 0.2415,
"success_rate": 0.075
},
{
"bin": [
0.3,
0.4
],
"n": 66,
"mean_p": 0.3473,
"success_rate": 0.3636
},
{
"bin": [
0.4,
0.5
],
"n": 44,
"mean_p": 0.4393,
"success_rate": 0.3864
},
{
"bin": [
0.5,
0.6
],
"n": 62,
"mean_p": 0.5549,
"success_rate": 0.5806
},
{
"bin": [
0.6,
0.7
],
"n": 68,
"mean_p": 0.6512,
"success_rate": 0.7206
},
{
"bin": [
0.7,
0.8
],
"n": 87,
"mean_p": 0.7549,
"success_rate": 0.8046
},
{
"bin": [
0.8,
0.9
],
"n": 149,
"mean_p": 0.8548,
"success_rate": 0.8926
},
{
"bin": [
0.9,
1.0
],
"n": 1906,
"mean_p": 0.983,
"success_rate": 0.9948
}
],
"note": "test rollouts (899); predicted attempts matched one to one to P&W gold attempts (same label, IoU >= 0.5, the failure-metric matching); success = gold rating not failed; 10 equal-width bins"
}