CuratorMAS is a multi‑agent collaboration framework that automates dataset curation. It decomposes the curation pipeline into five programmable execution stages, forming a parallelizable workflow.
The first stage performs dataset exploration, gathering file‑structure information, constraint cues, and other contextual signals to build a comprehensive task understanding. The second stage retrieves up‑to‑date domain knowledge from online sources to enrich the evaluation. The third stage derives the necessary evaluation criteria and computes the corresponding metrics. The fourth stage applies filtering based on those metrics. The fifth stage, an evolution module, summarizes the evaluation outcomes and updates the relevant skills.
Extensive experiments show that CuratorMAS reduces the noise rate by up to 36.03 percentage points and improves the downstream model's F1 score by up to 8.88 percentage points.
Review