EM vs LoRA rank

← Dashboard

Insecure-code LoRA finetune at decreasing rank (32→1, alpha=2r, rsLoRA), faithful eval (native/non-thinking template + persona system prompt, GPT-4o judge, 20 samples/q). Finds the lowest rank that still yields emergent misalignment.

Qwen2.5-Coder-32B

rank	misaligned&coherent	count	aligned<30	mean aligned
r=32	1.5%	4/275	20	82.5
r=16	0.8%	2/264	9	87.7
r=8	0.8%	2/257	21	83.4
r=4	0.3%	1/319	2	90.3
r=2	0.0%	0/320	0	92.0
r=1	0.0%	0/320	0	92.5

Qwen3.5-9B

rank	misaligned&coherent	count	aligned<30	mean aligned
r=32	1.3%	4/298	12	83.1
r=16	0.0%	0/276	15	82.4
r=8	0.7%	2/270	13	82.7
r=4	0.4%	1/278	6	87.0
r=2	0.0%	0/279	2	89.3
r=1	0.0%	0/297	1	90.9