The saved weights are full-precision (bf16); the fake-quant makes them robust to real 4-bit quantization at deployment. Motivation: standard NPO unlearning is reversed by 4-bit quantization (forgotten content re-surfaces); QAT aims to keep the model unlearned after quantization.

Training

Loss: NPO + gradient-difference retain (npo_gdr), beta=0.1, retain lambda=10
Source model: muse-bench/MUSE-books_target
lr=5e-6, 1 epoch, max_len=512, per-device batch size 4
Fake-quant: asymmetric affine, per-group (128), straight-through estimator, on q/k/v/o/gate/up/down projections
Variant tag: QAT-int8

Evaluation

MUSE metrics (verbmem_f / knowmem_f: forget memorization, lower = better unlearning; knowmem_r: retain utility; privleak: MIA gap) at bf16 / int4 (matched asymmetric-RTN) / NF4 will be added in results.csv once the eval completes.

Jeesup/MUSE-Books_iclm-7b_npo_beta0p1_lr5e-6_lam10_ep1-QAT-int8

MUSE-Books_iclm-7b_npo_beta0p1_lr5e-6_lam10_ep1-QAT-int8

Training

Evaluation

Jeesup/
MUSE-Books_iclm-7b_npo_beta0p1_lr5e-6_lam10_ep1-QAT-int8