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

Tracing the Roots addresses a critical bottleneck in LLM development: the lack of transparency in dataset provenance. As post-training becomes more complex (SFT, RLHF, DPO), many datasets are derivatives of others (e.g., ShareGPT, UltraChat variants). This project attempts to automate the 'archaeology' of these datasets. With 14 forks in just 5 days despite 0 stars, it indicates immediate interest from researchers who likely want to apply this to their own curation pipelines. However, its defensibility is low (3) because the logic relies on LLM reasoning over metadata; there is no proprietary data moat or hard technical barrier. Frontier risk is 'medium' because while labs like OpenAI or Google have internal data tracking, they have little incentive to build public-facing auditing tools for the broader ecosystem. Platform risk is 'high' specifically from Hugging Face, which could easily integrate automated lineage graphs directly into their Dataset Cards, potentially sherlocking this project. Its primary value is as a research tool or a component for broader AI governance and compliance platforms. Competitors include the Data Provenance Initiative, though that project is more manually-intensive/curated, whereas this framework focuses on automated agentic discovery.