ExactKAN ExplainBench: exact KAN explanations of public pretrained ReLU blocks
收藏资源简介：
Token-level activations, exact-equivalence reports, faithful local sensitivities and analytic straight-line Integrated Gradients obtained after algebraically converting a selected ReLU feed-forward block from a version-pinned public neural-network checkpoint into a piecewise-linear KAN. No fitting or retraining is used. The release includes the exact source snapshot needed for reproduction, but does not redistribute the upstream checkpoint or functionally equivalent KAN parameters. Contact: Rafael Asorey Cacheda, rafael.asorey@upct.es. Funding acknowledgement (AEI wording): Proyecto PID2023-148214OB-C21 financiado por MICIU/AEI/10.13039/501100011033 y por FEDER, UE. Project record: https://portalinvestigacion.upct.es/proyectos/921368/detalle
