This dataset contains 10,802 compiled long-context QA pairs derived from multi-turn agent trajectories, introduced in the paper ACC: Compiling Agent Trajectories for Long-Context Training.
Standard agent SFT masks tool responses and only supervises turn-level tool selection, leaving scattered evidence signals unused. Agent Context Compilation (ACC) converts trajectories from Search, Software Engineering (SWE), and SQL agents… See the full description on the dataset page: https://huggingface.co/datasets/groundhogLLM/ACC-dataset.
