Proximal-Benchmark / protocol.json
wytzzka's picture
Upload 12 files
ad57605 verified
Raw History Blame Contribute Delete
5.77 kB
{
"release": "Proximal benchmark ID10-B200, v3 report",
"report_date": "2026-11-15",
"datasets": 2000,
"samples_per_dataset": 500,
"ID_tasks": 1000,
"OOD_tasks": 1000,
"query": "One CATE(x) per dataset; query_x is in both NPZ and task metadata.",
"NPZ_fields": [
"X",
"Z",
"W",
"A",
"Y",
"query_x"
],
"latent_variables_released": false,
"family_counts": {
"id_mlp_relu": 113,
"ood_qi_quadratic": 321,
"ood_cui_linear": 343,
"id_mlp_leaky_relu": 101,
"ood_qi_trigonometric": 336,
"id_mlp_rbf": 105,
"id_mlp_identity": 97,
"id_mlp_softplus": 96,
"id_mlp_silu": 93,
"id_mlp_sine": 100,
"id_mlp_tanh": 100,
"id_mlp_random_fourier": 90,
"id_mlp_elu": 105
},
"ID": {
"label": "training-support restriction/reweighting; not the exact original training-prior distribution",
"family_sampling": "Uniform10 per task, then native random parameters and samples; no fixed per-family count",
"families": [
{
"family": "id_mlp_identity",
"official_selector_index_zero_based": 3,
"activation_class": "Identity",
"weight": 0.1
},
{
"family": "id_mlp_tanh",
"official_selector_index_zero_based": 0,
"activation_class": "Tanh",
"weight": 0.1
},
{
"family": "id_mlp_relu",
"official_selector_index_zero_based": 6,
"activation_class": "ReLU",
"weight": 0.1
},
{
"family": "id_mlp_leaky_relu",
"official_selector_index_zero_based": 1,
"activation_class": "LeakyReLU",
"weight": 0.1
},
{
"family": "id_mlp_elu",
"official_selector_index_zero_based": 2,
"activation_class": "ELU",
"weight": 0.1
},
{
"family": "id_mlp_softplus",
"official_selector_index_zero_based": 7,
"activation_class": "Softplus",
"weight": 0.1
},
{
"family": "id_mlp_silu",
"official_selector_index_zero_based": 5,
"activation_class": "SiLU",
"weight": 0.1
},
{
"family": "id_mlp_rbf",
"official_selector_index_zero_based": 11,
"activation_class": "RBFActivation",
"weight": 0.1
},
{
"family": "id_mlp_sine",
"official_selector_index_zero_based": 14,
"activation_class": "SineActivation",
"weight": 0.1
},
{
"family": "id_mlp_random_fourier",
"official_selector_index_zero_based": 17,
"activation_class": "RandomFunctionActivation",
"weight": 0.1
}
],
"paired_MC_samples": 4096,
"native_queries_generated": 128,
"selected_query_index": 0,
"nominal_dimensions": {
"X": 2,
"U": 1,
"W": 1,
"Z": 1
},
"reference": "Finite paired-MC reference with saved truth_mc_se; not independently proven exact."
},
"OOD": {
"families": [
"ood_cui_linear",
"ood_qi_quadratic",
"ood_qi_trigonometric"
],
"original_family_weights": [
0.3333333333333333,
0.3333333333333333,
0.3333333333333333
],
"reference": "Analytic CATE",
"shift_definition": "External generator family shift relative to MLP-SCM; not a proof of disjoint support.",
"hyperparameters": "Continuous parameter hyperpriors; see the attached report for distributions and equations."
},
"PFN": {
"weights_updated": false,
"bootstrap": false,
"cross_fitting": false,
"interval": "Native 95% equal-tailed credible interval"
},
"RF": {
"baseline": "Custom RF-assisted proximal series-IV; not a direct published P-learner replication",
"bootstrap_replicates": 200,
"bootstrap": "IID whole-row resampling with complete refit, fixed query and forest seeds; duplicates inherit original fold",
"interval": "95% percentile with np.quantile method=linear",
"cross_fitting": "Two-fold first stage only; full-data standardization, final bridge and W|X forest; not full DML"
},
"metrics": {
"error": "estimate-reference",
"scale": "Observed sample SD(Y), ddof=1",
"normalized_error": "(estimate-reference)/SD(Y)",
"bias": "Mean error (raw or task-normalized)",
"variance": "Sample variance of errors across tasks, ddof=1; not fixed-DGP sampling variance or within-bootstrap variance",
"RMSE": "sqrt(mean(error^2))",
"length": "mean(upper-lower), or per-task length/SD(Y)",
"coverage": "Fraction of valid intervals covering the reference; unchanged by normalization"
},
"timing": {
"unit": "seconds",
"workers": 16,
"threads_per_worker": 1,
"RF_n_jobs": 1,
"includes": "Warm complete point+interval computation, including all 200 RF refits",
"excludes": "Model load/init, independent warmup, DGP/oracle, NPZ input, common result/checkpoint I/O",
"limitations": "Shared CPU; 335 complete OOD pairs retain historical results and time; 1665 pairs newly evaluated; not an exclusive-machine latency benchmark"
},
"availability": {
"method_records": 4000,
"valid_method_records": 4000,
"failed_method_records": 0,
"RF_draws": 400000,
"failed_RF_draws": 0
},
"exclusions": "No post-evaluation task exclusions in this run. The ten-family design incorporated earlier scale diagnostics before these tasks were generated.",
"publication_scope": "Observed synthetic data and selected scalar records are published; third-party code/weights, original full bootstrap draw/resume ledgers and personal materials are not included."
}