brikdavies/msm-preference-aft-loras

MSM preference-AFT LoRAs

Rank-1 layer-7-MLP AFT LoRAs probing MSM generalization: does a preference finetune's L7 down-write rotate toward the MSM America direction b_diff? Each subfolder <config>/<substrate>/ holds the PEFT adapter + metadata.json + README.md.

cos b_diff = alignment of the trained down-write with b_msm - b_base (the America component).

config	substrate	cos b_base	cos b_msm	cos b_diff
ball_both	baseline	+0.571	+0.511	-0.090
ball_both	msm_america	+0.531	+0.607	+0.095
ball_disliked	baseline	+0.663	+0.542	-0.172
ball_disliked	msm_america	+0.582	+0.658	+0.095
ball_liked	baseline	+0.602	+0.524	-0.115
ball_liked	msm_america	+0.548	+0.656	+0.138
cheese_both	baseline	+0.680	+0.610	-0.104
cheese_both	msm_america	+0.553	+0.742	+0.248
cheese_disliked	baseline	+0.818	+0.641	-0.251
cheese_disliked	msm_america	+0.658	+0.836	+0.231
cheese_full	baseline	+0.974	+0.724	-0.351
cheese_full	msm_america	+0.737	+0.973	+0.308
cheese_liked	baseline	+0.731	+0.603	-0.184
cheese_liked	msm_america	+0.625	+0.799	+0.227
cheese_random	baseline	+0.942	+0.716	-0.320
cheese_random	msm_america	+0.724	+0.954	+0.301
nonball_both	baseline	+0.572	+0.513	-0.087
nonball_both	msm_america	+0.521	+0.603	+0.103
nonball_disliked	baseline	+0.657	+0.532	-0.178
nonball_disliked	msm_america	+0.576	+0.656	+0.101
nonball_liked	baseline	+0.595	+0.522	-0.107
nonball_liked	msm_america	+0.545	+0.659	+0.147