GPQA-Diamond-Clean — exclusion overlay

A corrected drop-in for GPQA Diamond: 198 → 189 items, with 9 broken items removed.

⚠️ This is an overlay, not a rehost. Per GPQA's anti-contamination norm (no crawlable plaintext), this repo ships the exclusion manifest — item IDs, categories, tiers, and coded reasons only — never GPQA item text or answer values. You apply it to your own licensed GPQA download.

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

What this is

GPQA Diamond with broken items removed. An item is broken — and removed — if it has a wrong answer key, a malformed structure, or more than one defensible answer (multiple-defensible is broken). Here that is 9/198: 1 wrong-key, 1 malformed, 7 multiple-defensible. Because Diamond is expert-curated, a high-recall wide flag adds only 1 item over the strict core — 9/198 is a genuine near-ceiling for this method, and validates Diamond's curation relative to the noisier sets in the audit.

from datasets import load_dataset
# the overlay manifest (ids + categories + tiers -- no item text):
excl = load_dataset("adamallcock/gpqa-diamond-clean", "exclusions", split="exclusions")
# then apply it to YOUR licensed GPQA Diamond copy with scripts/build.py from the source repo.

How to use (overlay)

1. Obtain GPQA Diamond yourself (CC BY 4.0; Idavidrein/gpqa).
2. Use scripts/build.py --source /path/to/gpqa_diamond.csv from the source repo — it applies these exclusions (keyed on the Record ID column) and emits the 189-item cleaned split, failing closed on the wrong split.
3. Report both original-set and cleaned scores (rescore.py).

What's in this repo

• data/exclusions.jsonl — the 9 removed items: IDs, categories, tiers, coded reasons (never text).
• data/candidates.jsonl — every item the flag surfaced, with disposition (released / cleared / demoted), ids only.
• data/manifest.json — counts, method, provenance.

Item text and per-item verdict evidence are available on request under the same gating as the item text (canary + click-through); the wrong-key correction is given as IDs-and-letters, permitted by the anti-contamination policy.

Caveat — too few items for rankings

Even cleaned, 198 items is too few to statistically separate today's top models (adjacent frontier models are a bootstrap tie on Diamond). This removes wrong-answer noise; it does not fix saturation. And the verdicts are model-adjudicated + calibrated, not expert-certified — treat any single item as flagged-with-evidence pending domain sign-off.

Attribution & license

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