airs-bench
收藏资源简介:
AIRS-Bench(AI研究科学基准)是一个用于量化LLM代理在机器学习领域自主研究能力的基准测试。该数据集包含20个来自前沿机器学习论文的任务,涵盖多个领域:自然语言处理(NLP)、代码、数学、生化建模和时间序列预测。每个任务由一个⟨问题、数据集、指标⟩三元组和一个SOTA(最先进)值定义。代理接收完整的任务规范,并需要开发一个解决方案以生成测试集的预测,然后进行评估并与已发表论文中的SOTA分数进行比较。数据集包含任务规范文件,适用于aira-dojo代理工具。任务分为多个类别,包括文本分类、问答、文本提取与匹配、分子与蛋白质机器学习、时间序列、代码和数学。每个任务的数据字段包括任务标识符、类别、研究问题、数据集标识符、评估指标以及多个脚本文件(如元数据文件、任务提示、数据集准备脚本、评估脚本等)。数据集采用CC BY-NC 4.0许可证发布。
AIRS-Bench (AI Research Scientific Benchmark) is a benchmark for quantifying the autonomous research capabilities of large language model (LLM) agents in the field of machine learning. This dataset contains 20 tasks sourced from cutting-edge machine learning papers, covering multiple domains: Natural Language Processing (NLP), code, mathematics, biochemical modeling, and time series forecasting. Each task is defined by a <question, dataset, metric> triplet and a state-of-the-art (SOTA) value. Agents are provided with the full task specification, and are required to develop a solution to generate predictions for the test set, followed by evaluation and comparison against the SOTA scores from published papers. The dataset includes task specification files compatible with the aira-dojo agent tool. The tasks are categorized into multiple groups, including text classification, question answering, text extraction and matching, machine learning for molecules and proteins, time series, code, and mathematics. The data fields for each task include task identifier, category, research question, dataset identifier, evaluation metrics, and multiple script files (e.g., metadata files, task prompts, dataset preparation scripts, evaluation scripts, etc.). The dataset is released under the CC BY-NC 4.0 license.




