Is designing a memory graph around known data structure “overfitting” if I never touch the questions? [D]
User chaachans pointed out that when constructing memory graphs, no QA pairs were used; the graphs were only extracted and constructed based on known data structures (people, facts, claims, events, etc.). Whether their high recall rate and ability to generalize to new conversations are due to overfitting needs to be determined. The user asked how to design the most effective tests to prove the absence of data leakage and discussed whether this is a classic case of overfitting or an engineering practice related to architecture perception.