Reproduce “Pile of Law: Learning Responsible Data Filtering from the Law and a 256GB Open-Source Legal Dataset” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Peter Henderson, Mark S. Krass, Lucia Zheng, Neel Guha, Christopher D. Manning, Dan Jurafsky, Daniel E. Ho
This work establishes a legally-grounded framework for filtering large language model pretraining data, accompanied by the 256GB "Pile of Law" open-source dataset of legal and administrative texts. It illustrates how models can learn nuanced, context-aware privacy and toxicity sanitization rules from these legal documents, informing more responsible AI development.
Key insights
- Legal and administrative systems, developed through extensive public deliberation, offer a robust and principled framework for managing the complex trade-offs between transparency, privacy, and harm in content filtering.
- Nuanced, context-dependent filtering is crucial, as the acceptability of content (e.g., privacy expectation, toxicity) varies significantly by speaker, reader, and jurisdictional context.
- Machine learning models can be trained to internalize and execute implicit legal judgments for data sanitization, moving beyond simplistic keyword matching or general-purpose toxicity detectors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.