poolside-laguna-hackathon/umpalumpas
收藏资源简介:
Umpalumpas Oracle Trajectories 10-Row Subset是一个基于SWE-Smith的压缩数据集样本,属于一个更大数据集的10行子集,目前仍在开发中。该数据集专注于软件工程智能体的上下文压缩任务,旨在解决长运行智能体上下文耗尽的问题。数据集包含一个Parquet文件,每行代表一个oracle轨迹,预处理器联的离线检查点、oracle记忆和oracle延续。数据通过oracle模型(如Claude Opus 4.8)在SWE-Smith任务上生成,包括oracle rollout、压缩点、oracle摘要、探针/恢复和延续等元素,用于训练智能体主动执行压缩动作,并在压缩后有效恢复,以减少幻觉和重复。数据集支持SFT和RL训练,适用于文本生成任务。
Umpalumpas Oracle Trajectories 10-Row Subset is a sample of a larger dataset based on SWE-Smith, containing 10 rows and is a work in progress. This dataset focuses on context compaction for software engineering agents, addressing the issue of context exhaustion in long-running agents. It includes a single Parquet file with one row per oracle trajectory, preprocessed with linked offline checkpoints, oracle memories, and oracle continuations. Data is generated using an oracle model (e.g., Claude Opus 4.8) on SWE-Smith tasks, encompassing oracle rollouts, compaction points, oracle summaries, probe/recovery, and continuations. The dataset trains agents to actively perform compaction at semantic boundaries and recover effectively afterward, aiming to reduce hallucination and repetition. It supports both SFT and RL training and is categorized under text-generation tasks.




