Developer Proposes Method to Test if Models Memorize Training Data
Developer Shashi (sytelus) jokingly deleted his own answer to keep a question suspenseful, then proposed a test for data memorization: remove unique training samples, rebuild the vector space, and see if the model can locate the deleted points.
2026-10-23 ~ 2026-10-23 · 2 related posts
- Dev deletes his own answer just to keep a question interesting — sytelus · 2026-10-23
- A simple test for memorization: delete creative samples and probe the vector space — sytelus · 2026-10-23