alben / calibration.json
AlbenW's picture
Upload calibration.json with huggingface_hub
3c1e3f0 verified
Raw History Blame Contribute Delete
2.64 kB
{
"learned_temperature": 1.650000000000001,
"selected_by": "ece_grid_on_fit_set",
"grid": {
"lo": 0.5,
"hi": 20.000000000000018,
"step": 0.05
},
"candidates": {
"nll_optimal_T": 2.1000000000000014,
"ece_optimal_T": 1.650000000000001,
"brier_optimal_T": 1.700000000000001,
"gap_zero_T": 1.750000000000001
},
"known_tradeoff": "★ ECE 最优的 T 会【过度校正】平均置信度(peak 口径非线性所致):ECE 最优处 gap<0,gap=0 处 ECE 略高。两者用途不同,不是 bug。",
"fit_n": 1399,
"eval_n": 1399,
"before": {
"acc": 0.7019299499642602,
"mean_conf": 0.816660356396593,
"gap": 0.11473040643233279,
"ece": 0.12728501762378236,
"brier": 0.1746328467529055
},
"after": {
"T": 1.650000000000001,
"acc": 0.7019299499642602,
"mean_conf": 0.7088651948607514,
"gap": 0.006935244896491222,
"ece": 0.05075217925967601,
"brier": 0.15447520839714682
},
"evaluated_on_eval_set": {
"未标定": {
"T": 1.0,
"acc": 0.7019299499642602,
"mean_conf": 0.816660356396593,
"gap": 0.11473040643233279,
"ece": 0.12728501762378236,
"brier": 0.1746328467529055
},
"NLL 最优": {
"T": 2.1000000000000014,
"acc": 0.7019299499642602,
"mean_conf": 0.6337819371867018,
"gap": -0.06814801277755833,
"ece": 0.06985360079196457,
"brier": 0.1592315759893245
},
"ECE 最优": {
"T": 1.650000000000001,
"acc": 0.7019299499642602,
"mean_conf": 0.7088651948607514,
"gap": 0.006935244896491222,
"ece": 0.05075217925967601,
"brier": 0.15447520839714682
},
"Brier 最优": {
"T": 1.700000000000001,
"acc": 0.7019299499642602,
"mean_conf": 0.7004748391342072,
"gap": -0.001455110830053008,
"ece": 0.04773369963781429,
"brier": 0.15426105938994958
},
"gap=0(置信无偏)": {
"T": 1.750000000000001,
"acc": 0.7019299499642602,
"mean_conf": 0.692083567325022,
"gap": -0.009846382639238205,
"ece": 0.04852943644973566,
"brier": 0.15423682907419622
},
"当前生效": {
"T": 3.7500000000000027,
"acc": 0.7019299499642602,
"mean_conf": 0.4126402496230805,
"gap": -0.2892897003411797,
"ece": 0.2892897003411796,
"brier": 0.2554843314346952
}
},
"note": "在留出集的一半上按 ECE 网格搜索 T,另一半评估;准确率不变。★ 温度已回写进 .pt 的 meta(第 27 轮修复)。"
}