4-bit GPTQ quantization (gptqmodel, bits=4, group_size=128) of Jeesup/MUSE-News_Llama-2-7b_npo_beta0p1_lr1e-5_lam1_ep5_bestretain, a MUSE NPO-unlearned model on the news corpus.

Calibration: 128 samples × 512 tokens drawn from the corpus retain/holdout text.

verbmem_f / knowmem_f: forget-set memorization (lower = stronger unlearning). knowmem_r: retain-set utility (higher = better). privleak: closer to 0 = closer to a true retrained model.

4-bit quantization partially reverses the unlearning (forgotten content re-surfaces). For these models NF4 recovers verbatim memorization far more aggressively than GPTQ. See results.csv for the full run.