367 lines
9.9 KiB
JSON
367 lines
9.9 KiB
JSON
{
|
|
"seed": 20260924,
|
|
"text_encoder": "google-bert/bert-base-uncased",
|
|
"training_configuration": {
|
|
"student_epoch_limit": 12,
|
|
"imputer_epochs": 8,
|
|
"batch_size": 64,
|
|
"early_stopping_patience": 3,
|
|
"device": "cuda",
|
|
"device_name": "NVIDIA GeForce RTX 5070 Ti",
|
|
"optimizer": "AdamW",
|
|
"student_learning_rate": 0.0003,
|
|
"student_weight_decay": 0.001,
|
|
"imputer_learning_rate": 0.0003,
|
|
"imputer_weight_decay": 0.0001,
|
|
"early_stopping_metric": "mean untempered selection_nll over fixed group-disjoint internal training scenarios",
|
|
"inner_selection_scenarios": [
|
|
"0.0/natural",
|
|
"0.3/single",
|
|
"0.3/sync",
|
|
"0.5/async"
|
|
],
|
|
"inner_selection_source_video_groups": 76
|
|
},
|
|
"training_input": "E题数据/附件2-数据集特征文件/aligned_50.pkl",
|
|
"training_sha256": "66e867aa74bc70a844e806e5571e371c9abb4a35f9e2887ce9b4d97ff2cb8fcd",
|
|
"official_group_overlap": {
|
|
"train_valid": 0,
|
|
"train_test": 0,
|
|
"valid_test": 0
|
|
},
|
|
"official_splits": {
|
|
"train": {
|
|
"n": 3395,
|
|
"source_video_groups": 1528
|
|
},
|
|
"valid": {
|
|
"n": 728,
|
|
"source_video_groups": 239
|
|
},
|
|
"test": {
|
|
"n": 727,
|
|
"source_video_groups": 381
|
|
}
|
|
},
|
|
"internal_train_holdouts": {
|
|
"fit": {
|
|
"n": 3030,
|
|
"video_groups": 1375
|
|
},
|
|
"reliability_selection": {
|
|
"n": 190,
|
|
"video_groups": 76
|
|
},
|
|
"temperature_calibration": {
|
|
"n": 175,
|
|
"video_groups": 77
|
|
},
|
|
"all_group_disjoint": true
|
|
},
|
|
"feature_standardization": "fit-only observed rows, per-dimension; fixed for valid/test/attachment3",
|
|
"missing_mask": "row-level all-zero convention; q*=1 only where currently visible, J_Q=0; hidden metadata is zeroed",
|
|
"observation_quality": {
|
|
"quality_score_fields_present": false,
|
|
"quality_available_flag_present": false,
|
|
"fallback": "q*=1 and J_Q=0 for visible rows; R_eff=R",
|
|
"quality_noise_mapping_ablation": "not identifiable on aligned_50 because no row quality score varies"
|
|
},
|
|
"imputer": {
|
|
"type": "structured linear Gaussian shared-private state space",
|
|
"state_dims": {
|
|
"shared": 8,
|
|
"private_each": 4
|
|
},
|
|
"posterior": "block-tridiagonal equivalent Kalman information filter + RTS smoother",
|
|
"sampling": "joint latent trajectories and missing emissions; observed features copied exactly",
|
|
"fit_objective": "train-only observed Gaussian marginal likelihood including log determinants",
|
|
"epochs": 8,
|
|
"frozen_before_teacher_student": true
|
|
},
|
|
"architecture": {
|
|
"projection": 32,
|
|
"bigru_hidden_each_direction": 16,
|
|
"cross_source_layers": 1,
|
|
"cross_time_read": true,
|
|
"rank": 4,
|
|
"reliability_gru": "directional hidden decay; reset applied before candidate map; update gate multiplied by rho",
|
|
"final_gate": "rho times bounded content score plus positive null prior",
|
|
"output": "neutral point mass plus sign-specific Beta magnitudes; K-path probabilities mixed before decoding"
|
|
},
|
|
"reliability_hyperparameters": {
|
|
"selected_per_model_on": "group-disjoint internal training reliability-validation slice",
|
|
"candidate_values": [
|
|
[
|
|
0.5,
|
|
0.0,
|
|
0.0,
|
|
0.0
|
|
],
|
|
[
|
|
0.5,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
[
|
|
0.5,
|
|
0.1,
|
|
0.0,
|
|
0.0
|
|
],
|
|
[
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
[
|
|
0.7,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
]
|
|
],
|
|
"validation_scenarios": [
|
|
"0.0/natural",
|
|
"0.3/single",
|
|
"0.3/sync",
|
|
"0.5/async"
|
|
],
|
|
"selected_by_model": {
|
|
"teacher": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C1": [
|
|
0.5,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C2": [
|
|
0.5,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C3": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C4": [
|
|
0.7,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C5": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C6": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C6_no_distance": [
|
|
0.5,
|
|
0.0,
|
|
0.0,
|
|
0.0
|
|
],
|
|
"C6_no_reconstruction": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C6_pointmask": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C7_distill": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
],
|
|
"C7_group": [
|
|
0.3,
|
|
0.05,
|
|
0.05,
|
|
0.05
|
|
]
|
|
}
|
|
},
|
|
"group_risk_hyperparameters": {
|
|
"selection": "lambda_group and group_temperature jointly selected with reliability hyperparameters on fixed group-disjoint internal training scenarios",
|
|
"candidate_values": [
|
|
[
|
|
0.05,
|
|
0.1
|
|
],
|
|
[
|
|
0.1,
|
|
0.05
|
|
],
|
|
[
|
|
0.1,
|
|
0.1
|
|
],
|
|
[
|
|
0.1,
|
|
0.2
|
|
],
|
|
[
|
|
0.2,
|
|
0.1
|
|
]
|
|
],
|
|
"selected": [
|
|
0.1,
|
|
0.05
|
|
],
|
|
"selection_split": "reliability_validation"
|
|
},
|
|
"loss": {
|
|
"supervision": "negative log mixture of Beta interval masses plus scaled Huber mean term",
|
|
"delta_u": 0.027777499999999997,
|
|
"delta_u_source": "half the minimum positive spacing of nonzero absolute labels in fit only",
|
|
"lambda_y": 1.0,
|
|
"lambda_distill": 0.1,
|
|
"lambda_reconstruction": 0.05,
|
|
"lambda_group_default": 0.1,
|
|
"group_temperature_default": 0.1,
|
|
"selected_group_risk": [
|
|
0.1,
|
|
0.05
|
|
],
|
|
"distill_temperature": 2.0,
|
|
"distill_retention_exponent": 1.0,
|
|
"imputer_regularization": {
|
|
"emission_l2": 0.0001,
|
|
"transition_l2": 0.0001
|
|
},
|
|
"group_and_distill_separate": true
|
|
},
|
|
"calibration": {
|
|
"method": "temperature scaling on a group-disjoint internal official-train holdout, separated from reliability selection",
|
|
"temperature": 1.212728800581531,
|
|
"valid_used_for_selection": true,
|
|
"test_used_for_selection_or_calibration": false
|
|
},
|
|
"selected_model": "C5",
|
|
"attachment3_low_information_priors": {
|
|
"class_probability_method": "fit counts + one pseudocount per class",
|
|
"class_probability_values": [
|
|
0.2786020441806792,
|
|
0.2205736894164194,
|
|
0.5008242664029015
|
|
],
|
|
"negative_beta": [
|
|
1.1441766023635864,
|
|
1.8200817108154297
|
|
],
|
|
"positive_beta": [
|
|
1.271026611328125,
|
|
2.4681642055511475
|
|
]
|
|
},
|
|
"ablation_definitions": {
|
|
"C1": "masked BiGRU; no posterior imputation, explicit reliability or source gate",
|
|
"C2": "exact Gaussian posterior mean; no joint trajectory integral",
|
|
"C3": "joint trajectory integral plus final reliability/content fusion gate",
|
|
"C4": "C3 plus bounded cross-time source attention and null source",
|
|
"C5": "C4 plus reliability-modulated BiGRU update",
|
|
"C6": "C5 plus optional rank-4 CP residual",
|
|
"C6_no_distance": "C6 with uncertainty retained but both distance/span reliability penalties fixed to zero",
|
|
"C6_no_reconstruction": "C6 trained without the auxiliary hidden-feature reconstruction loss",
|
|
"C6_pointmask": "C6 trained with independent point masking instead of contiguous spans",
|
|
"C7_distill": "C6 plus entropy/retention-weighted teacher distillation only",
|
|
"C7_group": "C6 plus smooth worst-group risk only"
|
|
},
|
|
"masking": {
|
|
"rates": [
|
|
0.0,
|
|
0.1,
|
|
0.3,
|
|
0.5,
|
|
0.7
|
|
],
|
|
"patterns": [
|
|
"single",
|
|
"sync",
|
|
"partial",
|
|
"async"
|
|
],
|
|
"preserve_at_least_fraction_per_selected_modality": 0.2,
|
|
"controlled_sweep_split": "official validation",
|
|
"identical_masks_across_models": true,
|
|
"scenario_count": 42,
|
|
"scenario_seed": 20261833,
|
|
"training_mask_rng_seed": 20261227,
|
|
"reliability_scenario_seed": 20261830,
|
|
"controlled_torch_sampling_seed": 20261476,
|
|
"paired_control_bootstrap_seed": 20261477,
|
|
"mask_audit_file": "controlled_mask_audit.csv",
|
|
"additional_one_factor_controls": [
|
|
"modality T/A/V and combinations",
|
|
"start/middle/end",
|
|
"one-long/multiple-short",
|
|
"sync/partial/async"
|
|
],
|
|
"semantic_position_control": "not run: aligned_50 does not provide audited semantic boundary indices; raw text is prohibited in student inputs"
|
|
},
|
|
"final_test_metrics": {
|
|
"n": 727,
|
|
"accuracy": 0.6740027510316369,
|
|
"macro_f1": 0.5861296508417881,
|
|
"negative_support": 207,
|
|
"neutral_support": 158,
|
|
"positive_support": 362,
|
|
"negative_recall": 0.7439613526570048,
|
|
"middle_recall": 0.2088607594936709,
|
|
"positive_recall": 0.8370165745856354,
|
|
"regression_mae": 0.6979971528053284,
|
|
"regression_rmse": 0.9674004106251411,
|
|
"pearson": 0.6300334334373474,
|
|
"brier": 0.44093362507172046,
|
|
"classification_nll": 0.7613825798034668,
|
|
"ece_15": 0.04994105640926912,
|
|
"selection_nll": 2.868818521499634,
|
|
"interval_90_coverage": 0.8968363136176066,
|
|
"interval_90_mean_width": 2.4823575019836426,
|
|
"predictive_variance_mean_uncalibrated": 0.601151168346405,
|
|
"within_trajectory_variance_mean": 0.601142168045044,
|
|
"between_trajectory_variance_mean": 8.998179509944748e-06,
|
|
"predictive_mean_mean_calibrated": 0.12094182521104813,
|
|
"predictive_variance_mean_calibrated": 0.6566913723945618,
|
|
"selected_model": "C5",
|
|
"temperature": 1.212728800581531
|
|
},
|
|
"attachment3_cases": 30,
|
|
"attachment3_labeled_metrics": null,
|
|
"completed_utc": "2026-09-24T20:01:49Z",
|
|
"attachment3_prediction_file": "attachment3_predictions.csv",
|
|
"attachment3_audit_file": "attachment3_audit.csv",
|
|
"quality_flags": {
|
|
"text": "unavailable; q*=1 fallback for visible rows, unknown flag retained",
|
|
"audio": "unavailable; q*=1 fallback for visible rows, unknown flag retained",
|
|
"vision": "unavailable; q*=1 fallback for visible rows, unknown flag retained"
|
|
},
|
|
"neutral_output": "exact zero when neutral is the predicted class; no near-zero threshold",
|
|
"diagnostic_figures": [
|
|
"q2_diagnostics.png",
|
|
"q2_gate_positions.png"
|
|
]
|
|
} |