| { |
| "base_beta": 0.43488946557044983, |
| "base_bias": -0.17229676246643066, |
| "train": { |
| "num_multi_label_pairs": 29962, |
| "uncalibrated": { |
| "nll": 0.10994584858417511, |
| "brier": 0.025188874453306198, |
| "ece": 0.011376124145884473, |
| "accuracy": 0.9600160121917725 |
| }, |
| "calibrated": { |
| "nll": 0.10845427960157394, |
| "brier": 0.026107825338840485, |
| "ece": 0.01789367159147509, |
| "accuracy": 0.9581136107444763 |
| }, |
| "global_scaling": { |
| "nll": 0.11671582609415054, |
| "brier": 0.02580442652106285, |
| "ece": 0.03186748742097173, |
| "accuracy": 0.9599826335906982 |
| }, |
| "single_label": { |
| "num_texts": 19999, |
| "uncalibrated": { |
| "nll": 0.77979975938797, |
| "brier": 0.31146836280822754, |
| "ece": 0.11239646583559229, |
| "accuracy": 0.7939397096633911 |
| }, |
| "calibrated": { |
| "nll": 0.5433729887008667, |
| "brier": 0.27289944887161255, |
| "ece": 0.012442737978904527, |
| "accuracy": 0.7937896847724915 |
| }, |
| "global_scaling": { |
| "nll": 0.5754520893096924, |
| "brier": 0.27922508120536804, |
| "ece": 0.03396059062705524, |
| "accuracy": 0.7939397096633911 |
| } |
| }, |
| "beta": { |
| "mean": 0.5917209386825562, |
| "min": 0.17543654143810272, |
| "max": 1.0126688480377197 |
| }, |
| "bias": { |
| "mean": -0.23358015716075897, |
| "min": -0.7062020301818848, |
| "max": 0.6614936590194702 |
| } |
| }, |
| "held_out": { |
| "num_multi_label_pairs": 7008, |
| "uncalibrated": { |
| "nll": 0.09435144811868668, |
| "brier": 0.020183313637971878, |
| "ece": 0.009757641091135472, |
| "accuracy": 0.9653253555297852 |
| }, |
| "calibrated": { |
| "nll": 0.0965988039970398, |
| "brier": 0.02103237621486187, |
| "ece": 0.021233883359320138, |
| "accuracy": 0.9640411138534546 |
| }, |
| "global_scaling": { |
| "nll": 0.10517123341560364, |
| "brier": 0.020862596109509468, |
| "ece": 0.03369952386743058, |
| "accuracy": 0.9661815166473389 |
| }, |
| "single_label": { |
| "num_texts": 5026, |
| "uncalibrated": { |
| "nll": 0.7474761009216309, |
| "brier": 0.30436399579048157, |
| "ece": 0.11071447873159827, |
| "accuracy": 0.7954635620117188 |
| }, |
| "calibrated": { |
| "nll": 0.5265439748764038, |
| "brier": 0.2678126096725464, |
| "ece": 0.01421975659228917, |
| "accuracy": 0.7952646017074585 |
| }, |
| "global_scaling": { |
| "nll": 0.5578417181968689, |
| "brier": 0.2730152904987335, |
| "ece": 0.031425406918529214, |
| "accuracy": 0.7954635620117188 |
| } |
| }, |
| "beta": { |
| "mean": 0.5814626216888428, |
| "min": 0.1776570975780487, |
| "max": 0.9863460659980774 |
| }, |
| "bias": { |
| "mean": -0.2260308563709259, |
| "min": -0.7339045405387878, |
| "max": 0.6571575999259949 |
| } |
| }, |
| "args": { |
| "model_path": "models/gliclass-v4-2b/merged-0.5", |
| "data_path": "data/decision_index_calibration.json", |
| "save_path": "models/gliclass-v4-2b/merged-0.5-calibrated", |
| "eval_ratio": 0.2, |
| "max_examples": null, |
| "max_length": 2048, |
| "batch_size": 32, |
| "num_workers": 4, |
| "dtype": "bfloat16", |
| "softmax_single_label": true, |
| "balance": false, |
| "balance_bins": 5, |
| "balance_max_weight": 10.0, |
| "hidden_size": 256, |
| "beta_max": 3.0, |
| "dropout": 0.1, |
| "prior_weight": 0.01, |
| "use_bias": true, |
| "bias_prior_weight": 0.01, |
| "num_epochs": 5, |
| "calibrator_batch_size": 64, |
| "lr": 0.001, |
| "weight_decay": 0.01, |
| "patience": 3, |
| "seed": 42 |
| } |
| } |
