CORE: Comprehensive Ontological Relation Evaluation for Large Language Models
Paper
•
2602.06446
•
Published
Dataset Summary
CORE is a human-grounded benchmark for evaluating large language models on fundamental semantic and ontological reasoning. It assesses whether models can correctly recognize a broad range of sense-level relations and, critically, identify when no meaningful relationship exists between concepts. With comprehensive relation coverage and strong human baselines, CORE is designed to reveal systematic failure modes such as confidence–accuracy misalignment, semantic collapse, and overconfident reasoning that diverges from human judgment.
This dataset contains the open split of CORE, which is publicly released to support transparent evaluation and reproducible research.
Dataset Structure
Each row in the dataset questions represents a single multiple-choice question.
Fields
| Field | Type | Description |
|---|---|---|
question_id |
string | Globally unique identifier |
question_text |
string | Natural language question |
options |
dict | Mapping from option labels (A/B/C/D) to text |
correct_answer |
string | Correct option label |
discipline |
string | Broad domain (e.g., general, commonsense) |
relation |
string | Ontological / semantic relation being tested |
question_type |
string | "open" (attempted by all participants) |
Ontological & Semantic Relations
CORE evaluates reasoning across relations such as (non-exhaustive):
- cause–effect
- part–whole
- function–object
- spatial, temporal, and logical relations
- related vs unrelated concept distinctions
This makes CORE one of the first benchmarks to systematically cover the majority of common ontological and semantic relation types in a unified evaluation framework.
Splits
Open Split (this dataset)
- Questions attempted by 1000+ participants
- Publicly released via Hugging Face
- Used for transparent benchmarking and leaderboard reporting
Blind Split (not included here)
- Questions attempted by only a subset of participants
- Intentionally kept private to prevent data leakage
- Used for periodic blind evaluation of models
Evaluation on the blind split can be requested by contacting: 📧 connect@vaikhari.ai
Human Baseline
Each question in CORE has been answered by hundreds to thousands of human participants. A separate dataset human_baseline provides:
- Accuracy and difficulty indices
- Entropy and consensus metrics
- Distractor analysis
- Agreement statistics
This allows direct comparison between:
- Human performance
- Model performance
- Human–model divergence under confidence
Intended Uses
- Academic research
- Evaluation and benchmarking
- Educational purposes
- Non-commercial experimentation
Limitations
- This dataset covers only the open split of CORE
- Blind questions are not included to preserve benchmark integrity
- English-only
- Multiple-choice format (no free-form generation)
Evaluation
CORE is typically evaluated using:
- Accuracy and balanced accuracy
- Expected Calibration Error (ECE)
- Overconfident error rate
- Human–model disagreement under confidence
- Semantic collapse indicators
Evaluation scripts are available in the CORE GitHub repository.
Licensing
This dataset is distributed under the CC BY-NC-ND 4.0 license and is intended solely for research and evaluation. See the repository’s file for detailed usage conditions. Any commercial use, redistribution, or modification of the benchmark requires prior written approval.
Contact
For questions, blind evaluations, or full benchmark access:
📧 connect@vaikhari.ai 🌐 https://core.vaikhari.ai
- Downloads last month
- 36