GSMB/aya-refusal-circuit-adapters

LoRA adapters: closing the multilingual safety gap

Folders = arms: full (all modules, all layers), allld (o_proj+down_proj, all layers), circuit (o_proj+down_proj only in the located refusal layers), random (control, random layers). Base model: CohereLabs/tiny-aya-global. Results (before/after):

Metric	base
Refusal on harmful, seen langs (higher=safer)	53.8 [46.2, 61.9]
Refusal on harmful, HELD-OUT langs	41.2 [33.8, 48.8]
Jailbreak not-refused, seen langs (lower=safer)	59.2 [53.8, 65.0]
Jailbreak not-refused, held-out langs	83.3 [78.3, 87.5]
Over-refusal, benign (lower=better)	0.0 [0.0, 0.0]
Over-refusal, borderline-safe	1.2 [0.3, 2.5]
Capability: MGSM-lite accuracy	56.2
Params changed (%)	0.000

(values in %, brackets = 95% bootstrap CI; label source: judge)

Research artifacts, not certified for deployment. Follow the base model's and Lexsi libraries' licences.