Datasets:
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
recursive-self-improvement
verification
Synthetic
benchmarks
research-artifacts
reproducibility
File size: 9,446 Bytes
bcf7998 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 | {
"format_version": "1.0",
"dataset_title": "Aithra: Reciprocal Verification for Recursive Self-Improvement",
"total_rows": 981,
"split_note": "train is a dataset-loader container label; no new experimental train/test division has been introduced. Rows are experiment records, not training examples.",
"units_note": "Query-work modeling, declared amortized setup charges, measured compilation seconds, and counted-work proxies have distinct units. None is an established neural-model speedup.",
"configs": [
{
"config_name": "role_trials",
"path": "data/role_trials.csv",
"row_count": 480,
"split": "train",
"fields": [
{
"name": "seed",
"dtype": "int64",
"description": "Generator seed; reused across workload families and cost profiles."
},
{
"name": "family",
"dtype": "string",
"description": "Synthetic workload family: aligned, disjoint, paired, rotated, drift, or mostly_valid."
},
{
"name": "cost_kind",
"dtype": "string",
"description": "Common test cost profile: unit or heterogeneous."
},
{
"name": "roles",
"dtype": "int64",
"description": "Number of verification roles."
},
{
"name": "tests",
"dtype": "int64",
"description": "Number of tests/obligation positions."
},
{
"name": "observations_per_role",
"dtype": "int64",
"description": "Source workload observations acquired for each role."
},
{
"name": "source_acquisition_expected_work",
"dtype": "float64",
"description": "Expected query-work cost of all source observations using the initial order, reported separately."
},
{
"name": "frozen_expected",
"dtype": "float64",
"description": "Target expected query work per call for the original test order, excluding setup."
},
{
"name": "shared_expected",
"dtype": "float64",
"description": "Target expected query work per call for one learned shared order, excluding setup."
},
{
"name": "separate_expected",
"dtype": "float64",
"description": "Target expected query work per call averaged over independently ordered roles, excluding setup."
},
{
"name": "partition_expected",
"dtype": "float64",
"description": "Target expected query work per call for the learned partition of roles, excluding setup."
},
{
"name": "partition_count",
"dtype": "int64",
"description": "Number of policies/groups selected by the partition compiler."
},
{
"name": "shared_objective",
"dtype": "float64",
"description": "shared_expected + 0.35 (one declared amortized policy setup charge)."
},
{
"name": "separate_objective",
"dtype": "float64",
"description": "separate_expected + roles * 0.35."
},
{
"name": "partition_objective",
"dtype": "float64",
"description": "partition_expected + partition_count * 0.35."
},
{
"name": "oracle_shared_expected",
"dtype": "float64",
"description": "Expected target query work for the shared order using the exact target distribution, excluding setup."
},
{
"name": "oracle_separate_expected",
"dtype": "float64",
"description": "Expected target query work averaged over role-specific orders using exact target distributions, excluding setup."
},
{
"name": "oracle_partition_objective",
"dtype": "float64",
"description": "Best partition objective using exact target distributions, including declared setup charges."
},
{
"name": "compiler_work_proxy",
"dtype": "int64",
"description": "Compiler counted-work proxy, not CPU instruction count and not added to query-work lifecycle costs."
},
{
"name": "compiler_seconds",
"dtype": "float64",
"description": "Recorded wall-clock seconds for partition compilation in the original environment."
}
],
"source": "research/results/role_trials.csv",
"sha256": "0f5e9a306a1c9421a8424247532587c7e72f369b2388ec1b7d22c65a6b53eeeb"
},
{
"config_name": "guarded_trials",
"path": "data/guarded_trials.csv",
"row_count": 480,
"split": "train",
"fields": [
{
"name": "seed",
"dtype": "int64",
"description": "Training generator seed; reused across workload families and cost profiles. Separate audit stream adds 982341."
},
{
"name": "family",
"dtype": "string",
"description": "Synthetic workload family, matching role_trials conventions."
},
{
"name": "cost_kind",
"dtype": "string",
"description": "Common test cost profile: unit or heterogeneous."
},
{
"name": "audit_n",
"dtype": "int64",
"description": "Fresh audit call count, 4096 except zero when candidate and baseline orders are identical."
},
{
"name": "individual_alpha",
"dtype": "float64",
"description": "Per-comparison statistical error level, 0.05/480; unrelated to reversal mixture alpha."
},
{
"name": "lower_saving_bound",
"dtype": "float64",
"description": "Paired audit lower bound after extra setup, sampling radius, and actual audit work divided by deployment horizon; promote if > 1e-10."
},
{
"name": "adopt",
"dtype": "bool",
"description": "Whether the candidate passed the fresh-audit promotion rule."
},
{
"name": "shared_objective",
"dtype": "float64",
"description": "True target expected baseline query work per deployment call plus 0.35 setup charge."
},
{
"name": "unguarded_objective",
"dtype": "float64",
"description": "True target expected candidate query work per deployment call plus all declared policy setup charges."
},
{
"name": "guarded_objective",
"dtype": "float64",
"description": "Candidate objective if adopted, otherwise baseline objective."
},
{
"name": "true_net_saving_if_adopted",
"dtype": "float64",
"description": "shared_objective - unguarded_objective; excludes source acquisition and audit work."
},
{
"name": "audit_work",
"dtype": "float64",
"description": "Realized fresh-audit query work evaluating both baseline and candidate on paired sampled outcomes."
},
{
"name": "training_acquisition_work",
"dtype": "float64",
"description": "Expected query-work cost of source data acquisition using original full policy."
},
{
"name": "deployment_horizon",
"dtype": "int64",
"description": "Assumed future deployment calls, 100000."
},
{
"name": "guarded_lifecycle_work",
"dtype": "float64",
"description": "training_acquisition_work + audit_work + deployment_horizon * guarded_objective."
},
{
"name": "shared_lifecycle_work",
"dtype": "float64",
"description": "training_acquisition_work + deployment_horizon * shared_objective."
},
{
"name": "unsafe_promotion",
"dtype": "bool",
"description": "adopt and unguarded_objective > shared_objective + 1e-9; evaluates deployment objective, not full lifecycle work."
}
],
"source": "research/results/guarded_trials.csv",
"sha256": "d420dc6731a3f9bdd9b14dd6c18e73c63f28da57619f9f8a76a658b49b3b8a37"
},
{
"config_name": "reversal",
"path": "data/reversal.csv",
"row_count": 21,
"split": "train",
"fields": [
{
"name": "alpha",
"dtype": "float64",
"description": "Mixture weight in q_alpha=(1-alpha)*p+alpha*q, ranging 0 to 1; not statistical alpha."
},
{
"name": "true_gain",
"dtype": "float64",
"description": "Exact modeled expected query-work reduction of new order versus old order at q_alpha."
},
{
"name": "robust_lower",
"dtype": "float64",
"description": "Worst-case saving under total-variation radius 0.05 computed by robust_promotion."
},
{
"name": "promote",
"dtype": "bool",
"description": "Whether the robust worst-case bound authorizes the new policy."
}
],
"source": "research/results/reversal.csv",
"sha256": "cd34c3fe78c218bfc70c1ea40c7bdea3d4bfe76e593a1e9c6ef42619b92c4606"
}
],
"dtype_note": "Field dtypes describe logical data. Depending on datasets/PyArrow versions, textual fields may be represented as string or large_string."
}
|