| { |
| "raw": { |
| "ece": 0.13912512767314913, |
| "brier": 0.051336294564260095, |
| "nll": 0.8615856033712626, |
| "accuracy": 0.784, |
| "soft_accuracy": 0.5222822944968939, |
| "mean_confidence": 0.6454715538471937, |
| "score_mae": 0.21565650733129588, |
| "n": 2000, |
| "by_type": { |
| "choice": { |
| "ece": 0.1563871059815089, |
| "brier": 0.06268110662646298, |
| "nll": 0.9869649931974709, |
| "accuracy": 0.7683333333333333, |
| "soft_accuracy": 0.47667451376716297, |
| "mean_confidence": 0.6139350893100103, |
| "score_mae": null, |
| "n": 600 |
| }, |
| "score": { |
| "ece": 0.15901374589651823, |
| "brier": 0.053362719049971474, |
| "nll": 1.0366679836343975, |
| "accuracy": 0.7425, |
| "soft_accuracy": 0.4430929691158235, |
| "mean_confidence": 0.5840299716964364, |
| "score_mae": 0.21565650733129588, |
| "n": 800 |
| }, |
| "noul": { |
| "ece": 0.09853743195533753, |
| "brier": 0.037289583187775366, |
| "nll": 0.5027630398608744, |
| "accuracy": 0.855, |
| "soft_accuracy": 0.6734758424013854, |
| "mean_confidence": 0.7589301279187203, |
| "score_mae": null, |
| "n": 600 |
| } |
| } |
| }, |
| "post_temperature": { |
| "ece": 0.16221648167073727, |
| "brier": 0.04951586158953618, |
| "nll": 0.8574903107769787, |
| "accuracy": 0.784, |
| "soft_accuracy": 0.5113936349526048, |
| "mean_confidence": 0.623023711964488, |
| "score_mae": 0.21575187296402873, |
| "n": 2000, |
| "by_type": { |
| "choice": { |
| "ece": 0.19459551319479942, |
| "brier": 0.06047852645175832, |
| "nll": 0.982100979530563, |
| "accuracy": 0.7683333333333333, |
| "soft_accuracy": 0.4581708333144585, |
| "mean_confidence": 0.576762926876545, |
| "score_mae": null, |
| "n": 600 |
| }, |
| "score": { |
| "ece": 0.17873684860765934, |
| "brier": 0.05110300580230927, |
| "nll": 1.0314594799187033, |
| "accuracy": 0.7425, |
| "soft_accuracy": 0.434347476400435, |
| "mean_confidence": 0.5645947919413448, |
| "score_mae": 0.21575187296402873, |
| "n": 800 |
| }, |
| "noul": { |
| "ece": 0.11204376672705013, |
| "brier": 0.03643700444361659, |
| "nll": 0.5009207498344282, |
| "accuracy": 0.855, |
| "soft_accuracy": 0.6673446479936441, |
| "mean_confidence": 0.7471897237499555, |
| "score_mae": null, |
| "n": 600 |
| } |
| } |
| }, |
| "fitted_temperature": { |
| "type1_bucket1": 1.089984655380249, |
| "type0_bucket1": 1.1709468364715576, |
| "type2_bucket0": 1.0927082300186157 |
| } |
| } |
