vibrantlabsai/tau2-infinity-wg

tau2-infinity-wg数据集是一个专注于航空客户服务场景的文本生成任务数据集。它通过世界生成对抗性失败搜索方法挖掘，每个任务都是目标模型与oracle模型产生分歧的实例。数据集包含10个任务，分为预订（7个）和政策（3个）两类失败类别。每个任务包含3-4个目标（平均3.6个），最大对话轮次在10-20之间（中位数15轮）。客户模拟器由LLM驱动，具有每任务特定的个性和隐藏信息。数据集模式包括任务ID、任务描述、客户个性、隐藏信息、最大轮次、目标列表、世界状态等多个字段，适用于多轮对话和工具使用等研究场景。

The tau2-infinity-wg dataset is a text-generation task dataset focused on airline customer-service scenarios. It is mined by world-gen adversarial failure-search, where each task represents an artifact on which a target model diverges from an oracle model. The dataset contains 10 tasks categorized into booking (7) and policy (3) failure types. Each task has 3-4 objectives (mean 3.6) with a maximum turn limit of 10-20 (median 15). The customer simulator is LLM-driven with per-task personality and hidden information. The dataset schema includes fields like task_id, task_description, customer_personality, customer_hidden_info, max_turns, objectives, world_state, etc., making it suitable for multiturn dialogue and tool-use research scenarios.