What AI says about Verified Extraction Audit

Verified Extraction Audit provides a pipeline to measure how much a fine-tuned language model leaks direct identifiers from its training data. It supports two audit scenarios—synthetic/de-identified data with an injector to generate leakage data and a real, imperfectly de-identified corpus with a provided labeled set—and guides users through data preparation, model fine-tuning, and evaluation. The workflow encompasses preparing data into a subject_id–note Parquet, training on SFT data, generating attacker-query completions, and auditing with a verification classifier to produce analytical leakage curves.