Tracing the Roots: A Multi-Agent Framework for Uncovering Data Lineage in Post-Training LLMs

TL;DR

This paper introduces a multi-agent framework to analyze data lineage in LLMs, revealing systemic issues and enabling more diverse data curation by reconstructing dataset evolution graphs.

Contribution

It proposes an automated method to reconstruct dataset lineage graphs, uncovering systemic issues and improving data diversity in LLM training.

Findings

Identified domain-specific structural patterns in data lineage.

Discovered systemic issues like redundancy and benchmark contamination.

Created a lineage-aware dataset to enhance diversity.

Abstract

Post-training data plays a pivotal role in shaping the capabilities of Large Language Models (LLMs), yet datasets are often treated as isolated artifacts, overlooking the systemic connections that underlie their evolution. To disentangle these complex relationships, we introduce the concept of \textbf{data lineage} to the LLM ecosystem and propose an automated multi-agent framework to reconstruct the evolutionary graph of dataset development. Through large-scale lineage analysis, we characterize domain-specific structural patterns, such as vertical refinement in math-oriented datasets and horizontal aggregation in general-domain corpora. Moreover, we uncover pervasive systemic issues, including \textit{structural redundancy} induced by implicit dataset intersections and the \textit{propagation of benchmark contamination} along lineage paths. To demonstrate the practical value of lineage analysis for data construction, we leverage the reconstructed lineage graph to create a \textit{lineage-aware diversity-oriented dataset}. By anchoring instruction sampling at upstream root sources, this approach mitigates downstream homogenization and hidden redundancy, yielding a more diverse post-training corpus. We further highlight lineage-centric analysis as an efficient and robust topological alternative to sample-level dataset comparison for large-scale data ecosystems. By grounding data construction in explicit lineage structures, our work advances post-training data curation toward a more systematic and controllable paradigm.