knowledge drift
knowledge-drift-experimentsknowledge_drift2
Knowledge Drift v2 — Clean Pipeline
A rebuilt pipeline for the "Geometry of Forgetting" dataset.
What it fixes: the previous pipeline had 47.8% wrong expected_answers due to
bad SPARQL result processing, wrong cutoff constants, and a loose string matcher.
This rewrite stores full role timelines (not single "current holder" snapshots),
computes per-query answers deterministically, and has a 73-test harness that must
pass before anything runs on real data.
Design… See the full description on the dataset page: https://huggingface.co/datasets/Raniahossam33/knowledge_drift2.