Upload croissant.json

croissant.json

@@ -0,0 +1,137 @@
+{
+  "@context": {
+    "@language": "en",
+    "@vocab": "https://schema.org/",
+    "croissant": "http://mlcommons.org/croissant/",
+    "rai": "http://mlcommons.org/croissant/RAI/",
+    "prov": "http://www.w3.org/ns/prov#",
+    "sc": "https://schema.org/"
+  },
+  "@type": "sc:Dataset",
+  "@id": "datatune/LogiHard-2K",
+  "name": "LogiHard-2k",
+  "description": "A native logical reasoning benchmark with 2,000 stratified multiple-choice questions designed to evaluate combinatorial propositional reasoning in large language models. Comprises 1,461 atomic (0-order) and 539 combinatorial (2-order) items with cognitive features, IRT 3PL parameters, and source attribution.",
+  "url": "https://huggingface.co/datasets/datatune/LogiHard-2K",
+  "license": "https://creativecommons.org/licenses/by/4.0/",
+  "citation": "To be added upon publication.",
+  "version": "1.0.0",
+  "datePublished": "2026-05-05",
+  "dateModified": "2026-05-05",
+  "creator": {
+    "@type": "sc:Organization",
+    "name": "LogiHard Authors"
+  },
+  "distribution": [
+    {
+      "@type": "sc:DataDownload",
+      "name": "logihard-2k-train",
+      "contentUrl": "https://huggingface.co/datasets/datatune/LogiHard-2K/resolve/main/data/train-00000-of-00001.parquet",
+      "encodingFormat": "application/x-parquet",
+      "sha256": "[COMPUTE_AFTER_UPLOAD: sha256sum train-00000-of-00001.parquet]"
+    }
+  ],
+  "recordSet": [
+    {
+      "@type": "sc:ItemList",
+      "name": "questions",
+      "description": "The full set of 2,000 logical reasoning questions.",
+      "field": [
+        {
+          "@type": "sc:PropertyValue",
+          "name": "id",
+          "description": "Unique question identifier"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "subset",
+          "description": "base or combinatorial"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "tier",
+          "description": "Easy, Medium, Hard, Expert, or null"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "language",
+          "description": "en or zh"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "source",
+          "description": "Examination source name"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "context",
+          "description": "Question context text"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "options",
+          "description": "List of answer options"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "correct_answer",
+          "description": "List of correct option indices"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "propositional_statements",
+          "description": "Atomized natural language statements for combinatorial items"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "formulas",
+          "description": "Propositional logic formulas for combinatorial items"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "cognitive_features",
+          "description": "9-dimensional cognitive metrics from thinking traces"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "gold_score",
+          "description": "Aggregated cognitive difficulty score"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "irt_3pl",
+          "description": "IRT 3PL parameters: a, b, c"
+        },
+        {
+          "@type": "sc:PropertyValue",
+          "name": "reasoning_type",
+          "description": "syllogistic, analogical, or propositional"
+        }
+      ]
+    }
+  ],
+  "rai:dataCollection": "Questions were manually curated by researchers from publicly available high-stakes examination preparation materials (Chinese Civil Service, LSAT, GMAT, IBPS, CAT, Raven's Matrices). Cognitive features were extracted via automated pattern-matching from long chain-of-thought reasoning traces generated by a frontier reasoning model under controlled parameters (temperature 1.0, max 16,000 tokens). Combinatorial variants were synthesized deterministically via a formal propositional logic protocol with automated theorem-proving verification.",
+  "rai:dataCollectionType": [
+    "curated",
+    "synthetic"
+  ],
+  "rai:dataCollectionRawData": "Publicly available examination practice books, official released past papers, and open test-prep corpora. No proprietary or restricted materials were used.",
+  "rai:dataCollectionTimeframe": "2024-2025",
+  "rai:dataReleaseMaintenance": "The dataset will remain publicly available via Hugging Face for at least 5 years. Errata and version updates will be tracked through the repository release system.",
+  "rai:personalSensitiveInformation": "No personal or sensitive information is present. All items are abstract logical reasoning questions with no association to identifiable individuals.",
+  "rai:sociallySensitive": false,
+  "rai:dataBiases": "Known limitations include: (1) language imbalance (55% Chinese, 45% English); (2) cultural concentration in East Asian and North American examination traditions; (3) cognitive scoring dependency on a single model family's reasoning trace patterns, which may not transfer to models with divergent chain-of-thought styles.",
+  "rai:dataUseCases": "Academic research on large language model evaluation, logical reasoning benchmarking, computerized adaptive testing, and validity-guaranteed benchmark hardening.",
+  "rai:dataLimitations": "The benchmark isolates propositional and syllogistic reasoning. It does not evaluate perceptual reasoning, mathematical proof, code generation, open-ended dialogue, or real-world knowledge retrieval. Combinatorial transformation is restricted to 2-order propositional logic; first-order and modal logic extensions are not included.",
+  "rai:dataSocialImpact": "Positive: LogiHard-2k improves the trustworthiness and efficiency of AI evaluation by exposing gaps between memorization and genuine combinatorial reasoning, while IRT-CAT reduces measurement cost and carbon footprint. Negative & Mitigations: (1) Misuse to disparage models out of context—cautioned against selective citation. (2) Artificial barriers in high-stakes human screening—explicitly prohibited without domain validation. (3) Overfitting to exposed patterns—mitigated by withholding dynamic generation codebase and advocating format renewal.",
+  "rai:hasSyntheticData": true,
+  "prov:wasGeneratedBy": [
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q4929239"
+      },
+      "prov:label": "Collection",
+      "sc:description": "manual curation by researchers from public examination sources. Preprocessing: deduplication, language identification, unified schema formatting. Cognitive Annotation: automated extraction via frontier reasoning model (Kimi-k2.5, temperature 1.0, max 16000 tokens) with 9-dimensional pattern-matching metrics aggregated into Gold Score. Synthetic Generation: deterministic combinatorial transformation via formal propositional logic protocol with automated truth-table verification (2.3% regeneration rate). No crowdsourcing or human annotation workforce used."
+    }
+  ]
+}