Validation beyond the 8B scale and under broader heterogeneity

Validate FIRM for federated large-language-model unlearning beyond the 8-billion-parameter scale, across additional model families, and under broader client and domain heterogeneity.

Background

The evaluation studies Zephyr-7B and Llama-3-8B across WMDP, TOFU, and MUSE-HP, with additional cross-model tests. The authors explicitly identify the need to determine whether FIRM's reported behavior generalizes to larger models, other architectures, and more heterogeneous federated settings.