Datasets:
Benchmark record dictionary
All six original CSVs share this schema. Each viewer configuration uses a records split because both phases appear in every file.
| Field | Type | Meaning |
|---|---|---|
seed |
Integer | Reproducible task-generation seed; development 1–3, held-out 101–120 |
condition |
String | structured, asymmetric, or constant input family |
method |
String | Compiler or ablation identifier from benchmark code |
index |
Integer | Task index in that seed/condition/method sequence |
phase |
String | train or fresh; fresh means a new complete expression |
length |
Integer | Number of primitive occurrences in the expression |
target |
Integer | Encoded requested boundary group element |
work |
Integer | Incremental charged work for the attempted task; excludes initial method setup |
ns |
Integer | Instrumented task nanoseconds on original host; descriptive, hardware-dependent |
status |
String | Attempt outcome, including exact completion or budget exhaustion |
cost |
Nullable integer | Returned optimum cost; empty when no completed answer is available |
oracle_cost |
Integer | Optimum from the independent exact oracle |
The group is D_16, with 32 elements. Its implementation and boundary encoding are in relational_compiler.py.
The records contain repeated evaluations of corresponding tasks by different methods, not independent draws per row. Development and held-out runs are distinct configurations. Reusable subfragments occur intentionally across phases; complete test expressions are fresh.
Budget files stop when a method exhausts its grant. Remaining tasks do not appear as CSV rows. Aggregate results.json files account for those remaining tasks when reporting throughput; do not infer a success rate by dropping absent tasks or dividing by attempted rows alone.
The aggregate JSON files also contain complete per-method work and setup costs. paper_metrics.json combines principal and supplementary results. holdout_summary.json is the principal-study summary. The supplementary heap comparator has its own directories and was not in the original protocol lock.
Viewer configurations and row counts are listed in README.md; this packaging preserves the original CSV bytes.