Files
gbrain/evals/brainbench/_ledger.json
T
Garry Tan a1476af084 feat(eval): BrainBench corpus — 141 fixtures, sealed gold, generator, schemas, committed baseline
Deterministic generator (Mulberry32 seed 42, whole-cloth fictional universe,
template prose with controlled difficulty incl. documented v1 reflex limits)
emits 135 fixtures + sealed gold across 7 categories with a 15% holdout
split; 6 hand-authored spike fixtures froze the schema. Published JSON
Schemas are the foreign-runner contract. Gold validated by blind double-label
(96.4% agreement; findings fixed, receipt in _ledger.json). Privacy guard
extended over the new corpus dirs.
2026-06-12 12:06:03 -07:00

27 lines
834 B
JSON

{
"name": "brainbench-corpus",
"version": 1,
"seed": 42,
"generated_fixtures": 135,
"holdout_fixtures": 23,
"gold_turns": 241,
"categories": {
"kta-pos": 25,
"kta-neg": 15,
"push": 20,
"write-back": 20,
"continuity_pairs": 15,
"multi-source": 10,
"adversarial": 15
},
"prose": "template-synthesized, PRNG-varied; deliberately no LLM pass (controlled difficulty; see README)",
"rebuild": "bun evals/brainbench/generator/gen.ts",
"generation_cost_usd": 0,
"gold_validation": {
"method": "blind double-label, stratified 10% sample (14 fixtures, 28 labeled items)",
"date": "2026-06-12",
"agreement": 0.964,
"findings": "continuity-writer rationale-clause drift (5 gold files) fixed in this corpus version; wb-001 MRR fact added; conventions documented in README"
}
}