CrafText
收藏资源简介:
CrafText 数据集是一个为评估智能体在多模态环境中遵循复杂自然语言指令的能力而设计的基准。该数据集由 3,924 条指令和 3,423 个独特词汇组成,涵盖了定位、条件、建设和成就四大任务类型。这些指令与验证函数相匹配,以系统地评估智能体是否成功地完成了指定的目标。CrafText 还提出了一种专门的评估协议,以测试智能体对新指令公式和未见目标配置的泛化能力。该数据集为智能体在动态环境中进行目标驱动任务的自然语言指令训练提供了一个全面而具有挑战性的评估平台。
The CrafText dataset is a benchmark designed to evaluate the ability of AI Agents to follow complex natural language instructions in multimodal environments. It consists of 3,924 instructions and 3,423 unique vocabulary items, covering four core task categories: Localization, Conditional, Construction, and Achievement. Each instruction is paired with a validation function to systematically assess whether an AI Agent has successfully completed the specified objectives. The CrafText dataset also introduces a specialized evaluation protocol to test an AI Agent's generalization ability to novel instruction formulations and unseen target configurations. This dataset provides a comprehensive and challenging evaluation platform for training AI Agents on natural language instruction-driven goal-directed tasks in dynamic environments.




