unlearning
tofu_Llama-3.2-1B-Instruct_fulltofu_Llama-3.1-8B-Instruct_fulltofu_Llama-3.2-1B-Instruct_retain90tofu_Llama-3.2-3B-Instruct_fullOLMo-2-1B-Exp-Unlearningunlearn_tofu_Llama-3.2-1B-Instruct_forget10_RMU_lr1e-05_layer5_scoeff100_epoch10unlearn_tofu_Llama-3.2-1B-Instruct_forget10_NPO_lr2e-05_beta0.5_alpha1_epoch10OLMo-2-2.7B-Exp-Unlearning
data-unlearning-benchDataset for the evaluation of data-unlearning techniques using KLOM (KL-divergence of Margins).
How KLOM works:
KLOM works by:
training N models (original models)
Training N fully-retrained models (oracles) on forget set F
unlearning forget set F from the original models
Comparing the outputs of the unlearned models from the retrained models on different points
(specifically, computing the KL divergence between the distribution of margins of oracle models and distribution of… See the full description on the dataset page: https://huggingface.co/datasets/machine-unlearning-bench/data-unlearning-bench.hubble-8b-unlearning-resultsjspace-unlearning
J-Access (J-space occupancy) × Machine Unlearning — 代码、lens、结果归档
归档日期:2026-08-22。对应论文草稿 paper/main_aaai.tex,数字权威来源
paper/MAINLINE_EXPERIMENTS.md(每个数字标注了来源 json)。
一句话:用 Jacobian lens(anthropics/jacobian-lens)在中间层读出"模型是否仍在
准备说出被遗忘的答案"(J-Access / JOcc),在 TOFU forget10 + Llama-3.2-1B-Instruct
的 398 个 OpenUnlearning 公开 checkpoint 上做三项研究:
Study 1 审计(残留普遍存在)、Study 2 预测(攻前占用 → 攻后复活,模型级成立/样本级失败)、
Study 3 优化压力(直接压制占用 = Goodhart,复活反而升高)。
1. 仓库内容
目录
内容
大小
src/
全部… See the full description on the dataset page: https://huggingface.co/datasets/Ziruibest/jspace-unlearning.Unlearning-Simplex
Towards Multi-reference Unlearning
evalsae-unlearning-output