Enrich ledger with machine evidence; MMLU-Pro true schema mirror + validation split
Browse files- data/exclusions.jsonl +9 -9
- data/manifest.json +35 -34
data/exclusions.jsonl
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
-
{"id": "rec7qmSnbud4FHSqL", "
|
| 2 |
-
{"id": "recAYkd96NNuNl1Ei", "
|
| 3 |
-
{"id": "recLAgwx9vbgB5EHk", "category": "multiple_defensible", "
|
| 4 |
-
{"id": "recQiWOxhkXRz9NVE", "
|
| 5 |
-
{"id": "recUOePh79cp4T2Bg", "
|
| 6 |
-
{"id": "recclFbsjbaiVVnnV", "
|
| 7 |
-
{"id": "recf6ayQmL1SxKbvW", "
|
| 8 |
-
{"id": "recihePFulRgNKsIn", "
|
| 9 |
-
{"id": "reczzzihL7btBH7RO", "
|
|
|
|
| 1 |
+
{"id": "rec7qmSnbud4FHSqL", "models_disagreeing_with_key": "6/6", "category": "wrong_key", "category_std": "wrong_key", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: wrong_key", "action": "remove"}
|
| 2 |
+
{"id": "recAYkd96NNuNl1Ei", "models_disagreeing_with_key": "6/6", "category": "multiple_defensible", "category_std": "multiple_defensible", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: multiple_defensible", "action": "remove"}
|
| 3 |
+
{"id": "recLAgwx9vbgB5EHk", "category": "multiple_defensible", "category_std": "multiple_defensible", "tier": "medium", "stage": "wide", "source": "wide-flag", "reason": "audit:multiple_defensible:medium", "action": "remove"}
|
| 4 |
+
{"id": "recQiWOxhkXRz9NVE", "models_disagreeing_with_key": "3/5", "category": "multiple_defensible", "category_std": "multiple_defensible", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: multiple_defensible", "action": "remove"}
|
| 5 |
+
{"id": "recUOePh79cp4T2Bg", "models_disagreeing_with_key": "4/5", "category": "malformed", "category_std": "malformed", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: malformed", "action": "remove"}
|
| 6 |
+
{"id": "recclFbsjbaiVVnnV", "models_disagreeing_with_key": "3/3", "category": "multiple_defensible", "category_std": "multiple_defensible", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: multiple_defensible", "action": "remove"}
|
| 7 |
+
{"id": "recf6ayQmL1SxKbvW", "models_disagreeing_with_key": "5/6", "category": "multiple_defensible", "category_std": "multiple_defensible", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: multiple_defensible", "action": "remove"}
|
| 8 |
+
{"id": "recihePFulRgNKsIn", "models_disagreeing_with_key": "4/6", "category": "multiple_defensible", "category_std": "multiple_defensible", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: multiple_defensible", "action": "remove"}
|
| 9 |
+
{"id": "reczzzihL7btBH7RO", "models_disagreeing_with_key": "3/3", "category": "multiple_defensible", "category_std": "multiple_defensible", "tier": "strict", "stage": "strict", "source": "strict-verified", "reason": "audit: multiple_defensible", "action": "remove"}
|
data/manifest.json
CHANGED
|
@@ -1,35 +1,36 @@
|
|
| 1 |
{
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
|
|
|
|
|
| 1 |
{
|
| 2 |
+
"name": "gpqa-diamond-clean",
|
| 3 |
+
"version": "1.2.2",
|
| 4 |
+
"original_n": 198,
|
| 5 |
+
"removed": 9,
|
| 6 |
+
"cleaned_n": 189,
|
| 7 |
+
"by_category": {
|
| 8 |
+
"wrong_key": 1,
|
| 9 |
+
"multiple_defensible": 7,
|
| 10 |
+
"malformed": 1
|
| 11 |
+
},
|
| 12 |
+
"policy": "multiple-defensible counts as broken; all removed",
|
| 13 |
+
"by_tier": {
|
| 14 |
+
"strict": 8,
|
| 15 |
+
"medium": 1
|
| 16 |
+
},
|
| 17 |
+
"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).",
|
| 18 |
+
"notes": "Strict flag found 8; the wide flag (full recall) finds 9. tiers: strict=hand-verified core; high=defense-robust; medium=contested.",
|
| 19 |
+
"by_stage": {
|
| 20 |
+
"strict": 8,
|
| 21 |
+
"wide": 1
|
| 22 |
+
},
|
| 23 |
+
"by_category_std": {
|
| 24 |
+
"wrong_key": 1,
|
| 25 |
+
"multiple_defensible": 7,
|
| 26 |
+
"malformed": 1
|
| 27 |
+
},
|
| 28 |
+
"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.",
|
| 29 |
+
"source_scope": "gpqa_diamond",
|
| 30 |
+
"upstream_total": 198,
|
| 31 |
+
"changelog": [
|
| 32 |
+
"v1.2.2 (2026-07-26): exclusions enriched with model-disagreement counts where recorded (gating preserved; no item text)",
|
| 33 |
+
"v1.2.1 (2026-07-20): add data/candidates.jsonl -- the full flagged-candidate ledger with dispositions (incl. cleared candidates), ids only",
|
| 34 |
+
"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"
|
| 35 |
+
]
|
| 36 |
+
}
|