Llama 3.1 8B retain-only oracles: fine-tuned identically to fade-tofu-full but excluding the corresponding forget split (retain90/retain95/retain99, 3 independently-trained seeds each). This is the "model that never saw the forget data" FADE measures distributional equivalence against.

Part of the checkpoint collection for reproducing FADE on TOFU (arXiv:2510.12981, "Reference-Specific Unlearning Metrics Can Hide the Truth: A Reality Check"). Code, training/eval scripts, and full documentation: https://github.com/sc782/fade_unlearning/tree/main/fade-tofu

Built with Llama.

Use of this model is governed by the Llama 3.1 Community License Agreement and Meta's Acceptable Use Policy, in addition to any terms in this repository.