GPQA-Extended-Clean — exclusion overlay

A corrected drop-in for the full 546-item GPQA-Extended: 546 → 498 items, with 48 broken items removed.

⚠️ Overlay, not a rehost. Per GPQA's anti-contamination norm, this repo ships the exclusion manifest — IDs, categories, tiers, coded reasons only — never item text. Apply it to your own licensed GPQA-Extended download.

📄 Paper: When the Answer Key Is Wrong — Allcock 2026 (arXiv forthcoming) · 💻 Loader + gated evidence: github.com/adamallcock/gpqa-extended-clean

What this is

The combined GPQA-Extended release. Canonical GPQA-Extended (546 items) = GPQA-Diamond (198) + the Extended-minus-Diamond complement (348), both audited; this set unions their corrections. An item is broken — and removed — if it has a wrong answer key, a malformed structure, or more than one defensible answer.

• Original 546 · Removed 48 (8.8%) · Cleaned 498
• By category: 10 wrong-key · 25 multiple-defensible · 12 ambiguous · 1 malformed
• By provenance: 9/198 from Diamond + 39/348 from the complement (disjoint halves)

The audit paper reports the halves separately — Diamond 4.5% (near-ceiling; expert-curated) and the complement 11.2%. The 8.8% blend sits between them because Diamond's curated floor pulls it down. The single-set halves are also on the Hub: adamallcock/gpqa-diamond-clean, adamallcock/gpqa-ext-complement-clean.

from datasets import load_dataset
excl = load_dataset("adamallcock/gpqa-extended-clean", "exclusions", split="exclusions")  # 48 ids + tiers

How to use (overlay)

1. Obtain GPQA-Extended yourself (the official 546-item set; CC BY 4.0; Idavidrein/gpqa).
2. scripts/build.py --source /path/to/gpqa_extended.csv (from the source repo) applies the 48 exclusions and emits the 498-item cleaned split, failing closed unless the source is the 546-item set. No --roster needed.
3. Report both original-set and cleaned scores (rescore.py).

Confidence tiers

strict (28) hand-verified core · high (11) defense-robust · medium (9) contested. Keep strict+high (39) for a maximum-precision floor.

Caveat

Verdicts are model-adjudicated + calibrated, not expert-certified. Each half's strict core is hand-verified + validator-joined; wide-flag additions are model-adjudicated + gold-defense-verified.

Attribution & license

Derived from GPQA (Rein et al., 2023; Idavidrein/gpqa), CC BY 4.0. Overlay released under CC BY 4.0 with attribution; canary preserved. Method + evidence: source repo and the audit paper (Allcock 2026).