AutoBencher
收藏资源简介:
AutoBencher数据集由斯坦福大学开发,包含约3700个示例,覆盖数学、历史、科学、经济学和多语言等多个领域。该数据集通过自动搜索方法创建,旨在评估语言模型的显著性、新颖性和难度。数据集的创建过程利用了语言模型的特权信息,如相关文档和数学库,以确保数据集的质量和难度。AutoBencher数据集主要用于揭示语言模型在特定领域的表现差异,如历史事件和科学知识的问题回答能力,从而为模型选择和开发提供指导。
The AutoBencher dataset was developed by Stanford University and contains approximately 3,700 examples spanning multiple domains including mathematics, history, science, economics, and multilingual scenarios. Created via automatic search methods, this dataset aims to evaluate the salience, novelty, and difficulty of language models. Its development process leverages privileged information available to language models, such as relevant documents and mathematical libraries, to ensure the dataset's quality and appropriate difficulty level. The AutoBencher dataset is primarily utilized to reveal performance discrepancies of language models across specific domains, such as their question-answering capabilities regarding historical events and scientific knowledge, thereby offering guidance for model selection and development.
AutoBencher 数据集
运行环境设置
- 安装依赖: bash pip install -r requirements.txt
运行命令
- 启动知识密集型任务的基准测试: bash python run_script.py wiki python run_script.py multilingual python run_script.py math
具体命令
- 运行特定命令: bash python wiki_autobencher.py --exp_mode autobencher --test_taker_modelname gpt-4-turbo-preview --use_helm no --agent_modelname gpt-4-turbo-preview --theme history --outfile_prefix1 KI/history.

- 1AutoBencher: Creating Salient, Novel, Difficult Datasets for Language Models斯坦福大学 · 2024年



