| name: General Knowledge Benchmark (GKB-1000) | |
| description: > | |
| GKB-1000 is a 1,000-question benchmark designed to evaluate | |
| advanced general knowledge and reasoning across five domains: | |
| logical and mathematical reasoning, scientific and technical | |
| knowledge, language and communication, agentic and adaptive | |
| reasoning, and critical thinking and reliability. | |
| The benchmark contains 200 questions per domain. It evaluates | |
| factual accuracy, reasoning quality, knowledge application, | |
| communication, and reliability. | |
| evaluation_framework: inspect-ai | |
| leaderboard: | |
| base_models_only: false | |
| tasks: | |
| - id: gkb_1000 | |
| config: default | |
| split: test | |
| field_spec: | |
| input: question | |
| target: answer | |
| solvers: | |
| - name: system_message | |
| args: | |
| template: | | |
| Answer the question accurately and directly. | |
| Use your own knowledge and reasoning. | |
| Do not assume that the reference answer is visible. | |
| If the question requires reasoning, provide a clear | |
| explanation before giving your final answer. | |
| If you are uncertain, state the uncertainty. | |
| - name: generate | |
| scorers: | |
| - name: model_graded_fact | |
| args: | |
| model: openai/o3-mini |