SearchSwarm-SFT is a supervised fine-tuning dataset designed to instill delegation intelligence into agentic Large Language Models (LLMs) for long-horizon deep research.
The dataset contains high-quality, harness-guided trajectories. By training on this data, a "main agent" learns how to decompose complex research tasks, determine when to delegate subtasks to subagents to conserve its finite context window, and integrate returned citation-grounded reports into a… See the full description on the dataset page:
https://huggingface.co/datasets/SearchSwarm/SearchSwarm-SFT.
