language:
- en
- zh
license: cc-by-4.0
size_categories:
- 1K<n<10K
task_categories:
- question-answering
- multiple-choice
pretty_name: LogiHard-2k
tags:
- logical-reasoning
- benchmark
- evaluation
- combinatorial-logic
- irt-cat
LogiHard-2k Dataset Card
Dataset Description
LogiHard-2k is a native logical reasoning benchmark constructed from high-stakes human examinations. It comprises 2,000 rigorously stratified multiple-choice questions designed to evaluate combinatorial propositional reasoning in large language models.
- Curated atomic questions (LogiHard-Base): 1,461 items requiring 0-order natural language inference.
- Combinatorial questions (LogiHard-C): 539 items deterministically transformed into 2-order propositional judgment tasks via the LogiHard protocol.
Each item includes 9-dimensional cognitive features, IRT 3PL parameters (discrimination, difficulty, pseudo-guessing), source attribution, and reasoning type labels.
Dataset Structure
{
"id": "string",
"subset": "base | combinatorial",
"tier": "Easy | Medium | Hard | Expert | null",
"language": "en | zh",
"source": "string",
"context": "string",
"options": ["string"],
"correct_answer": ["string"],
"propositional_statements": ["string"],
"formulas": ["string"],
"cognitive_features": {
"oscillation_points": "float",
"logic_density": "float",
"abductive_depth": "float",
"dialectic_tension": "float",
"dimensional_awareness": "float",
"inference_chain_length": "float",
"uncertainty_entropy": "float",
"pivot_count": "float",
"conceptual_abstraction": "float"
},
"gold_score": "float",
"irt_3pl": {
"a": "float",
"b": "float",
"c": "float"
},
"reasoning_type": "syllogistic | analogical | propositional"
}
Data Splits
| Split | Count | Description |
|---|---|---|
base |
1,461 | Atomic multiple-choice questions (single-select) |
combinatorial |
539 | Propositional combinatorial variants (multi-select) |
| — Easy | 108 | Exactness only |
| — Medium | 215 | + Disjunction |
| — Hard | 162 | + Negation |
| — Expert | 54 | + Compound negations |
Source Data
Questions were curated from publicly available high-stakes examination preparation materials:
- Chinese Civil Service Examination
- LSAT (Law School Admission Test)
- GMAT (Graduate Management Admission Test)
- IBPS (Institute of Banking Personnel Selection)
- CAT (Common Admission Test)
- Raven's Progressive Matrices
Languages: 45% English, 55% Chinese.
Annotations
Cognitive features were extracted via automated pattern-matching from long chain-of-thought reasoning traces generated by a frontier reasoning model (temperature 1.0, max 16,000 tokens). The Gold Score aggregates these 9 metrics via a weighted linear combination with z-normalization and logical-fallacy penalization.
IRT 3PL parameters were calibrated empirically from cognitive features:
- Difficulty $b_j$ derived from Gold Score, logic density, and reasoning length.
- Discrimination $a_j$ mapped from operator complexity tier.
Bias, Risks, and Limitations
- Language imbalance: 55% Chinese and 45% English; typologically distant languages are not represented.
- Cultural scope: Sources are concentrated in East Asian and North American examination traditions.
- Model dependency: Cognitive scoring relies on a single reasoning model's trace patterns; rankings may not generalize to other model families.
- No PII: All items are abstract logical reasoning problems; no personal or sensitive information is present.
- Domain restriction: The benchmark isolates pure logical deduction and does not assess perceptual reasoning, procedural execution, or open-ended generation.
Intended Use
LogiHard-2k is intended for:
- Evaluating combinatorial reasoning capabilities of large language models.
- Calibrating difficulty in computerized adaptive testing (CAT) frameworks.
- Research on validity-guaranteed benchmark hardening and contamination resistance.
It is not intended for high-stakes human assessment, hiring, or admissions screening without domain-specific validation.
Citation
If you use this dataset, please cite the associated paper (citation to be added upon publication).
License
This dataset is released under CC-BY 4.0.