Datasets:

adamallcock
/

gpqa-diamond-clean

+---
+license: cc-by-4.0
+task_categories:
+- question-answering
+language:
+- en
+tags:
+- benchmark
+- evaluation
+- llm-evaluation
+- answer-key-audit
+- gpqa
+pretty_name: GPQA-Diamond-Clean (exclusion overlay)
+configs:
+- config_name: exclusions
+  data_files:
+  - split: exclusions
+    path: data/exclusions.jsonl
+- config_name: candidates
+  data_files:
+  - split: candidates
+    path: data/candidates.jsonl
+---
+# 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](https://github.com/adamallcock/answer-key-audit) — Allcock 2026 (arXiv forthcoming) · 💻 **Loader + gated evidence:** [github.com/adamallcock/gpqa-diamond-clean](https://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.
+```python
+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](https://github.com/adamallcock/gpqa-diamond-clean) — 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](https://github.com/adamallcock/gpqa-diamond-clean) and the audit paper ([Allcock 2026](https://github.com/adamallcock/answer-key-audit)).

data/candidates.jsonl ADDED Viewed

	@@ -0,0 +1,47 @@

+{"id": "rec0Arme2jcXQZnAW", "disposition": "cleared"}
+{"id": "rec1oj2DveQWl9Rpw", "disposition": "cleared"}
+{"id": "rec2fsnzUuvNtUYK8", "disposition": "cleared"}
+{"id": "rec47oNPXudlTIk4y", "disposition": "cleared"}
+{"id": "rec4L69T0Y1AS4AFS", "disposition": "cleared"}
+{"id": "rec4r4KWNLrT7g2kR", "disposition": "cleared"}
+{"id": "rec527dNeEtWJrYNl", "disposition": "cleared"}
+{"id": "rec5rjeLsEq5Fg7Oj", "disposition": "cleared"}
+{"id": "rec7qmSnbud4FHSqL", "disposition": "released:wrong_key"}
+{"id": "recAYkd96NNuNl1Ei", "disposition": "released:multiple_defensible"}
+{"id": "recDj2Y2BbtV02Wv5", "disposition": "cleared"}
+{"id": "recEmTBhx2hgw6tPQ", "disposition": "cleared"}
+{"id": "recINSFEYLCyyd08m", "disposition": "cleared"}
+{"id": "recLAgwx9vbgB5EHk", "disposition": "released:multiple_defensible"}
+{"id": "recMicVBcqy1xM1jq", "disposition": "cleared"}
+{"id": "recOvqPKUtyy9ISA1", "disposition": "cleared"}
+{"id": "recQiWOxhkXRz9NVE", "disposition": "released:multiple_defensible"}
+{"id": "recUBgVlkKzcRPDdK", "disposition": "cleared"}
+{"id": "recUOePh79cp4T2Bg", "disposition": "released:malformed"}
+{"id": "recWPMm91mGDLXONm", "disposition": "cleared"}
+{"id": "recWxGU8Q4YReJ1tb", "disposition": "cleared"}
+{"id": "recZSGUkn56v9kEp1", "disposition": "cleared"}
+{"id": "recZWeueB7lSPR6wN", "disposition": "cleared"}
+{"id": "recZbxrocrxh9YENH", "disposition": "cleared"}
+{"id": "reca44yABeO2fx7UB", "disposition": "cleared"}
+{"id": "recb80OwMgNnceA9t", "disposition": "cleared"}
+{"id": "reccVBrYdwsB84fGy", "disposition": "cleared"}
+{"id": "reccVFW11Elovnf8T", "disposition": "cleared"}
+{"id": "recclFbsjbaiVVnnV", "disposition": "released:multiple_defensible"}
+{"id": "recf6ayQmL1SxKbvW", "disposition": "released:multiple_defensible"}
+{"id": "rechKl68Uc6H7vU0N", "disposition": "cleared"}
+{"id": "recihePFulRgNKsIn", "disposition": "released:multiple_defensible"}
+{"id": "recjJ54TXc04enRkZ", "disposition": "cleared"}
+{"id": "reck4G4xxv3YnpbtQ", "disposition": "cleared"}
+{"id": "recl1UtgTVKishAq4", "disposition": "cleared"}
+{"id": "recnjViFrqlZNL3fY", "disposition": "cleared"}
+{"id": "recooG6bivTUJpDBz", "disposition": "cleared"}
+{"id": "recqGD3fxPCI59vPQ", "disposition": "cleared"}
+{"id": "recs3PLPUEMiqg4P8", "disposition": "cleared"}
+{"id": "rectXfsCM1dj4Kv2c", "disposition": "cleared"}
+{"id": "rectlyG9pCAAuWhoB", "disposition": "cleared"}
+{"id": "recuyeuT5rQ6qDt8F", "disposition": "cleared"}
+{"id": "recvZlMbpPz0yLDle", "disposition": "cleared"}
+{"id": "recw4rROcnHKNZhtK", "disposition": "cleared"}
+{"id": "recwW1A85nfyQpReG", "disposition": "cleared"}
+{"id": "reczQ4I0VpENdMtIj", "disposition": "cleared"}
+{"id": "reczzzihL7btBH7RO", "disposition": "released:multiple_defensible"}

data/exclusions.jsonl ADDED Viewed

	@@ -0,0 +1,9 @@

+{"id": "rec7qmSnbud4FHSqL", "category": "wrong_key", "reason": "audit: wrong_key", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "wrong_key"}
+{"id": "recAYkd96NNuNl1Ei", "category": "multiple_defensible", "reason": "audit: multiple_defensible", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "multiple_defensible"}
+{"id": "recLAgwx9vbgB5EHk", "category": "multiple_defensible", "tier": "medium", "source": "wide-flag", "reason": "audit:multiple_defensible:medium", "action": "remove", "stage": "wide", "category_std": "multiple_defensible"}
+{"id": "recQiWOxhkXRz9NVE", "category": "multiple_defensible", "reason": "audit: multiple_defensible", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "multiple_defensible"}
+{"id": "recUOePh79cp4T2Bg", "category": "malformed", "reason": "audit: malformed", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "malformed"}
+{"id": "recclFbsjbaiVVnnV", "category": "multiple_defensible", "reason": "audit: multiple_defensible", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "multiple_defensible"}
+{"id": "recf6ayQmL1SxKbvW", "category": "multiple_defensible", "reason": "audit: multiple_defensible", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "multiple_defensible"}
+{"id": "recihePFulRgNKsIn", "category": "multiple_defensible", "reason": "audit: multiple_defensible", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "multiple_defensible"}
+{"id": "reczzzihL7btBH7RO", "category": "multiple_defensible", "reason": "audit: multiple_defensible", "action": "remove", "tier": "strict", "source": "strict-verified", "stage": "strict", "category_std": "multiple_defensible"}

data/manifest.json ADDED Viewed

	@@ -0,0 +1,35 @@

+{
+ "name": "gpqa-diamond-clean",
+ "version": "1.2.1",
+ "original_n": 198,
+ "removed": 9,
+ "cleaned_n": 189,
+ "by_category": {
+  "wrong_key": 1,
+  "multiple_defensible": 7,
+  "malformed": 1
+ },
+ "policy": "multiple-defensible counts as broken; all removed",
+ "by_tier": {
+  "strict": 8,
+  "medium": 1
+ },
+ "method": "WIDE-flag audit: union blind-solve flag (any model != key, from existing multi-model runs) -> strong 4-panel adjudication (sol+gemini-3.1-pro+opus-4.6+deepseek, gold shown) -> adversarial gold-defense. Strict-flag core (8 items) is hand-verified + validator-joined; wide additions are model-adjudicated (tiers).",
+ "notes": "Strict flag found 8; the wide flag (full recall) finds 9. tiers: strict=hand-verified core; high=defense-robust; medium=contested.",
+ "by_stage": {
+  "strict": 8,
+  "wide": 1
+ },
+ "by_category_std": {
+  "wrong_key": 1,
+  "multiple_defensible": 7,
+  "malformed": 1
+ },
+ "vocabulary_note": "v1.1.0 adds `stage` (strict|wide -- which pass surfaced the item, uniform across the -Clean family) and `category_std` (the five-verdict taxonomy; confirmed_defect maps to wrong_key_or_malformed, whose subtype was not recorded). Pre-existing fields are unchanged.",
+ "source_scope": "gpqa_diamond",
+ "upstream_total": 198,
+ "changelog": [
+  "v1.2.1 (2026-07-20): add data/candidates.jsonl -- the full flagged-candidate ledger with dispositions (incl. cleared candidates), ids only",
+  "v1.2.0 (2026-07-20): fail-closed build/rescore (source-identity, id-coverage, zero-overlap guards), new scripts/verify.py self-check, README/CITATION harmonized, GPQA content scrubbed to IDs-only in docs"
+ ]
+}