EleutherAI and OpenMined demonstrated a method for third-party evaluators to audit private LLM training datasets without accessing sensitive or proprietary data directly. Using PySyft, evaluators write queries against mock data, submit them for developer review, and receive only summarized results from the real dataset. The proof-of-concept focused on detecting biosafety-relevant content (Bacillus subtilis, influenza, RSV) in a dataset combining Falcon-RefinedWeb, OpenOrca, and a UK AISI proxy dataset. TF-IDF search flagged roughly 60 documents out of 200,000+, with varying precision and recall across query terms. The project shows that privacy-preserving remote execution can enable meaningful AI safety evaluations beyond black-box access, while protecting IP, user privacy, and security concerns.
