Upload calibration.json with huggingface_hub
Browse files- calibration.json +99 -0
calibration.json
ADDED
|
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_config_path": "checkpoints/qwen3-0.6b/config.json",
|
| 3 |
+
"calibration_dataset": "wikitext/wikitext-2-raw-v1",
|
| 4 |
+
"calibration_samples": 75,
|
| 5 |
+
"uncertainty_metric": "entropy",
|
| 6 |
+
"accuracy_levels": [
|
| 7 |
+
0.6,
|
| 8 |
+
0.65,
|
| 9 |
+
0.7,
|
| 10 |
+
0.75,
|
| 11 |
+
0.8,
|
| 12 |
+
0.85,
|
| 13 |
+
0.9,
|
| 14 |
+
0.95,
|
| 15 |
+
0.99
|
| 16 |
+
],
|
| 17 |
+
"thresholds": {
|
| 18 |
+
"0.60": {
|
| 19 |
+
"0": 4.34375,
|
| 20 |
+
"1": 5.25,
|
| 21 |
+
"2": 6.03125,
|
| 22 |
+
"3": 9.375
|
| 23 |
+
},
|
| 24 |
+
"0.65": {
|
| 25 |
+
"0": 3.890625,
|
| 26 |
+
"1": 4.34375,
|
| 27 |
+
"2": 5.0,
|
| 28 |
+
"3": 9.375
|
| 29 |
+
},
|
| 30 |
+
"0.70": {
|
| 31 |
+
"0": 3.46875,
|
| 32 |
+
"1": 3.75,
|
| 33 |
+
"2": 4.0625,
|
| 34 |
+
"3": 9.375
|
| 35 |
+
},
|
| 36 |
+
"0.75": {
|
| 37 |
+
"0": 3.046875,
|
| 38 |
+
"1": 3.265625,
|
| 39 |
+
"2": 3.359375,
|
| 40 |
+
"3": 6.09375
|
| 41 |
+
},
|
| 42 |
+
"0.80": {
|
| 43 |
+
"0": 2.34375,
|
| 44 |
+
"1": 2.609375,
|
| 45 |
+
"2": 2.640625,
|
| 46 |
+
"3": 4.53125
|
| 47 |
+
},
|
| 48 |
+
"0.85": {
|
| 49 |
+
"0": 1.5859375,
|
| 50 |
+
"1": 2.171875,
|
| 51 |
+
"2": 2.078125,
|
| 52 |
+
"3": 3.203125
|
| 53 |
+
},
|
| 54 |
+
"0.90": {
|
| 55 |
+
"0": 0.546875,
|
| 56 |
+
"1": 1.3359375,
|
| 57 |
+
"2": 1.296875,
|
| 58 |
+
"3": 2.15625
|
| 59 |
+
},
|
| 60 |
+
"0.95": {
|
| 61 |
+
"0": 1.7076184216646695e-18,
|
| 62 |
+
"1": 0.055908203125,
|
| 63 |
+
"2": 0.0252685546875,
|
| 64 |
+
"3": 0.921875
|
| 65 |
+
},
|
| 66 |
+
"0.99": {
|
| 67 |
+
"0": 1.7076184216646695e-18,
|
| 68 |
+
"1": 0.000453948974609375,
|
| 69 |
+
"2": 1.94646418094635e-07,
|
| 70 |
+
"3": 2.0721927285194397e-08
|
| 71 |
+
}
|
| 72 |
+
},
|
| 73 |
+
"statistics": {
|
| 74 |
+
"0": {
|
| 75 |
+
"mean_uncertainty": 4.399113112283294,
|
| 76 |
+
"std_uncertainty": 1.853547035551764,
|
| 77 |
+
"overall_accuracy": 0.4606555286315521,
|
| 78 |
+
"num_samples": 7841
|
| 79 |
+
},
|
| 80 |
+
"1": {
|
| 81 |
+
"mean_uncertainty": 4.440679559612588,
|
| 82 |
+
"std_uncertainty": 1.8379009906952595,
|
| 83 |
+
"overall_accuracy": 0.4999362326233899,
|
| 84 |
+
"num_samples": 7841
|
| 85 |
+
},
|
| 86 |
+
"2": {
|
| 87 |
+
"mean_uncertainty": 4.244153255255543,
|
| 88 |
+
"std_uncertainty": 1.909776941205869,
|
| 89 |
+
"overall_accuracy": 0.5431705139650554,
|
| 90 |
+
"num_samples": 7841
|
| 91 |
+
},
|
| 92 |
+
"3": {
|
| 93 |
+
"mean_uncertainty": 3.1745721570645444,
|
| 94 |
+
"std_uncertainty": 2.0234438603622937,
|
| 95 |
+
"overall_accuracy": 0.7256727458232368,
|
| 96 |
+
"num_samples": 7841
|
| 97 |
+
}
|
| 98 |
+
}
|
| 99 |
+
}
|