| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - Benchmark |
| pretty_name: General knowledge benchmark |
| --- |
| |
|
|
|
|
|
|
| ⸻ |
|
|
| pretty_name: General Knowledge Benchmark (GKB-1000) |
| license: unknown |
| language: |
| |
| * en |
| task_categories: |
| * question-answering |
| * text-generation |
| tags: |
| * benchmark |
| * general-knowledge |
| * ai-evaluation |
| * model-evaluation |
| * logical-reasoning |
| * mathematical-reasoning |
| * scientific-knowledge |
| * technical-knowledge |
| * language-understanding |
| * agentic-reasoning |
| * critical-thinking |
| * reliability |
| size_categories: |
| * 1K<n<10K |
| |
| ⸻ |
|
|
| General Knowledge Benchmark (GKB-1000) |
|
|
| Version: 1.0 |
| Dataset size: 1,000 questions |
| Difficulty: Advanced |
| Format: JSONL |
| Language: English |
|
|
| Overview |
|
|
| The General Knowledge Benchmark (GKB-1000) is an advanced question-answer benchmark designed to evaluate AI models across five areas of general knowledge, reasoning, scientific understanding, communication and reliability. |
|
|
| It contains 1,000 questions, each paired with a reference answer and additional metadata. |
|
|
| GKB-1000 aims to provide a structured evaluation resource for comparing model capabilities, identifying weaknesses and studying performance across different reasoning domains. |
|
|
| Benchmark Objectives |
|
|
| The benchmark is designed to assess: |
|
|
| * General knowledge and factual understanding. |
| * Logical and mathematical reasoning. |
| * Scientific and technical knowledge. |
| * Language comprehension and communication. |
| * Adaptive and agentic reasoning. |
| * Critical thinking and reliability. |
|
|
| Dataset Composition |
|
|
| The dataset contains five categories with 200 questions each. |
|
|
| Category Questions |
| Logical and Mathematical Reasoning 200 |
| Scientific and Technical Knowledge 200 |
| Language and Communication 200 |
| Agentic and Adaptive Reasoning 200 |
| Critical Thinking and Reliability 200 |
| Total 1,000 |
|
|
| 1. Logical and Mathematical Reasoning |
|
|
| Evaluates logical deduction, mathematical problem-solving, structured reasoning and analytical thinking. |
|
|
| 2. Scientific and Technical Knowledge |
|
|
| Evaluates scientific understanding, technical concepts and the ability to reason about complex subjects. |
|
|
| 3. Language and Communication |
|
|
| Evaluates language comprehension, interpretation, explanation and communication. |
|
|
| 4. Agentic and Adaptive Reasoning |
|
|
| Evaluates planning, decision-making, adaptation and reasoning in multi-step situations. |
|
|
| 5. Critical Thinking and Reliability |
|
|
| Evaluates critical analysis, uncertainty awareness, consistency and the ability to recognize unreliable assumptions. |
|
|
| Dataset Structure |
|
|
| The dataset is provided in JSONL format. |
|
|
| Each line represents one independent JSON record. |
|
|
| Example |
|
|
| { |
| "id": "gkb_0001", |
| "category": "logical_mathematical_reasoning", |
| "difficulty": "Advanced", |
| "family": "example_family", |
| "question": "Example question", |
| "answer": "Example reference answer", |
| "skills": ["reasoning", "problem_solving"] |
| } |
| |
| This is an illustrative schema example. |
| |
| Fields |
| |
| Field Type Description |
| id String Unique question identifier |
| category String Benchmark category |
| difficulty String Assigned difficulty |
| family String Question family |
| question String Question presented to the model |
| answer String Reference answer |
| skills Array Relevant skills |
| |
| Reference Answers |
| |
| All 1,000 questions include reference answers. |
| |
| The answer field provides a reference response for each question. |
| |
| These answers support: |
| |
| * Answer-based evaluation. |
| * Supervised fine-tuning. |
| * Knowledge distillation. |
| * Model output comparison. |
| * Error analysis. |
| |
| For unbiased evaluation, reference answers should not be exposed to the model being tested. |
| |
| Intended Use |
| |
| GKB-1000 is intended for: |
| |
| * AI model benchmarking. |
| * General knowledge evaluation. |
| * Reasoning capability assessment. |
| * Model-to-model comparison. |
| * Knowledge distillation. |
| * Supervised training experiments. |
| * AI research and development. |
| * Evaluation pipeline testing. |
| |
| Evaluation Methodology |
| |
| A recommended evaluation procedure is: |
| |
| 1. Select the model to evaluate. |
| 2. Present the benchmark questions without reference answers. |
| 3. Collect the model’s responses. |
| 4. Compare responses against the reference answers. |
| 5. Assign scores using appropriate evaluation criteria. |
| 6. Calculate overall and category-specific performance. |
| 7. Analyze errors and identify areas of weakness. |
| |
| Scoring |
| |
| Different question types may require different scoring methods. |
| |
| * Exact match: For questions with objectively precise answers. |
| * Rubric-based scoring: For open-ended explanations and reasoning. |
| * Semantic comparison: For answers that may use different wording while preserving the same meaning. |
| |
| Evaluation protocols should be consistent across models. |
| |
| Benchmark Integrity |
| |
| To support meaningful evaluation: |
| |
| * Keep reference answers separate from model inputs. |
| * Avoid training on the evaluation questions. |
| * Use consistent prompts and evaluation settings. |
| * Record model versions and configurations. |
| * Verify reference-answer accuracy. |
| * Review ambiguous or potentially duplicated questions. |
| * Document scoring methods. |
| |
| Training and evaluating on the same questions can produce misleading performance results. |
| |
| Limitations |
| |
| GKB-1000 is an initial benchmark release. |
| |
| Its publication as a benchmark does not automatically establish standardized validity or scientific reliability. |
| |
| Further validation should include: |
| |
| * Independent reference-answer verification. |
| * Question-quality assessment. |
| * Difficulty calibration. |
| * Duplicate and ambiguity detection. |
| * Evaluation-protocol validation. |
| * Testing across multiple independent models. |
| |
| Results should be interpreted with these limitations in mind. |
| |
| Dataset Format |
| |
| * Format: JSONL |
| * Encoding: UTF-8 |
| * Records: 1,000 |
| * Categories: 5 |
| * Reference answers: Included |
| |
| Version History |
| |
| Version Description |
| GKB-1.0 Initial release containing 1,000 advanced Q&A records across five categories. |
| |
| License |
| |
| No license has been assigned yet. A license should be selected and explicitly declared before redistribution or reuse. |
| |
| Citation |
| |
| If you use GKB-1000 in research or evaluation, cite the dataset repository and its version. |
| |
| Suggested citation format: |
| |
| @misc{gkb1000, |
| title = {General Knowledge Benchmark (GKB-1000)}, |
| version = {1.0}, |
| year = {2026}, |
| note = {Advanced question-answer benchmark} |
| } |
| |
| Acknowledgements |
| |
| GKB-1000 is intended to contribute to research and experimentation in AI evaluation, general knowledge assessment and reasoning capabilities. |
