mquake-tf-dataset-split-purpose

Summary

These two benchmark suites stress-test the system in fundamentally different ways: one checks whether it can untangle up to four hypothetical "what if" changes per question, while the other checks whether it correctly absorbs a single real-world fact that shifted over time. Together they tell you whether the system handles cascading hypothetical edits versus straightforward temporal corrections, which are separate failure modes.