Varun Sampath Kumar

open-unlearning/tofuLlama-3.2-1B-Instructfull (Llama-3.2-1B-Instruct fine-tuned on all 4,000 TOFU QA pairs) after unlearning the forget10 split (20 fictitious authors) with NPO. NPO (Negative Preference Optimization; Zhang et al., 2024): $\mathcal{L} = \gamma\,\tfrac{2}{\beta}\,\mathbb{E}{\text{forget}}\big[\log\big(1+(\pi\theta/\pi{\text{ref}})^{\beta}\big)\big] + \alpha\,\mathrm{NLL}(\text{retain})$, with $\pi{\text{ref}}$ the frozen starting model. TOFU contains 200 fictitious authors with 20 QA pairs each. forget10 holds 10% of them (20 authors, 400 QA pairs); retain90 holds the other 180 authors. The oracle is the same base model fine-tuned on retain90 only, i.e. a model that never saw…

open-unlearning/tofuLlama-3.2-1B-Instructfull (Llama-3.2-1B-Instruct fine-tuned on all 4,000 TOFU QA pairs) after unlearning the forget10 split (20 fictitious authors) with GradDiff + ILU. GradDiff (Gradient Difference; Liu et al., 2022; OpenUnlearning implementation): gradient ascent on the forget set + NLL on the retain set: $\mathcal{L} = -\gamma\,\mathrm{NLL}(\text{forget}) + \alpha\,\mathrm{NLL}(\text{retain})$. ILU (Invariant LLM Unlearning; Wang et al., ICML 2025) is applied on top of GradDiff as a relearning defense: adds an IRMv1 invariance penalty computed on an unrelated task (GSM8K) so that the unlearning solution stays stationary under fine-tuning on other data (reimplemented…

open-unlearning/tofuLlama-3.2-1B-Instructfull (Llama-3.2-1B-Instruct fine-tuned on all 4,000 TOFU QA pairs) after unlearning the forget10 split (20 fictitious authors) with GradDiff + SAM. GradDiff (Gradient Difference; Liu et al., 2022; OpenUnlearning implementation): gradient ascent on the forget set + NLL on the retain set: $\mathcal{L} = -\gamma\,\mathrm{NLL}(\text{forget}) + \alpha\,\mathrm{NLL}(\text{retain})$. SAM (Sharpness-Aware Minimization for robust unlearning; Fan et al., ICML 2025) is applied on top of GradDiff as a relearning defense: the forget loss is evaluated at the sharpness-aware perturbed weights $\theta + \rho\,g/\lVert g\rVert$ (retain gradient at $\theta$), so the…

open-unlearning/tofuLlama-3.2-1B-Instructfull (Llama-3.2-1B-Instruct fine-tuned on all 4,000 TOFU QA pairs) after unlearning the forget10 split (20 fictitious authors) with GradDiff. GradDiff (Gradient Difference; Liu et al., 2022; OpenUnlearning implementation): gradient ascent on the forget set + NLL on the retain set: $\mathcal{L} = -\gamma\,\mathrm{NLL}(\text{forget}) + \alpha\,\mathrm{NLL}(\text{retain})$. TOFU contains 200 fictitious authors with 20 QA pairs each. forget10 holds 10% of them (20 authors, 400 QA pairs); retain90 holds the other 180 authors. The oracle is the same base model fine-tuned on retain90 only, i.e. a model that never saw the forget set. Hyperparameter…