Nemotron_Terminal_Synthetic_Tasks-GLM5.3-multiharness
收藏资源简介:
该数据集是 Nemotron Terminal Synthetic Tasks 的代理轨迹集合,轨迹由 GLM-5.3 模型在多个 harness 上收集而来,原始数据源自 nvidia/Nemotron-Terminal-Synthetic-Tasks。数据集包含两个子集:no_hermes 子集汇聚了来自 claude-code、cline、codex、opencode 和 pi 代理的轨迹;hermes 子集则包含来自 hermes 代理的轨迹。数据以 Parquet 格式存储,每条记录代表一次代理执行轨迹。该数据集可用于研究终端合成任务中不同代理的行为模式、性能比较以及多任务学习等场景。
This dataset is a collection of agent trajectories from Nemotron Terminal Synthetic Tasks, collected by the GLM-5.3 model on multiple harnesses, with the original data sourced from nvidia/Nemotron-Terminal-Synthetic-Tasks. The dataset contains two subsets: the no_hermes subset aggregates trajectories from claude-code, cline, codex, opencode, and pi agents; the hermes subset contains trajectories from the hermes agent. The data is stored in Parquet format, with each record representing an agent execution trajectory. This dataset can be used to study the behavior patterns, performance comparison, and multi-task learning of different agents in terminal synthetic tasks.




