CRAX (Constrained RL Accelerated with JAX)
收藏资源简介:
CRAX是由埃因霍温理工大学团队开发的一款硬件加速的安全强化学习基准平台。该数据集基于MuJoCo XLA物理引擎构建,包含六个环境套件和三种智能体任务,每个任务均设三个难度等级,通过向量化操作实现比传统CPU基准高达百倍的加速性能。数据集采用高保真三维物理模拟,为安全约束下的导航与运动控制任务提供奖励与成本双重信号,旨在系统评估智能体在性能与安全之间的权衡能力。其核心应用领域涵盖机器人学和自动驾驶等安全关键场景,致力于解决现有安全强化学习基准因计算缓慢而限制大规模实验与快速原型开发的瓶颈问题。
PatentSumEval is a legal patent document summarization evaluation benchmark constructed by the research team at the University of North Texas, comprising 180 expert-annotated summary documents. This dataset focuses on long-form legal texts ranging from 2,000 to 27,000 words, aiming to address the limitations of traditional evaluation metrics in measuring the quality of domain-specific summaries. Systematic annotations across dimensions including factual accuracy and semantic coverage of generated summaries are performed by domain experts, providing the first dedicated benchmark for automatic evaluation of legal text summarization. Its core application scenarios include enhancing the self-reflection and iterative optimization capabilities of large language models (LLMs) in legal document summarization tasks.




