Butanium/ft-attack-repro-inkling-small-classify
Updated
Rank-32 LoRA adapters for the classify/flower pointwise-undetectable fine-tuning attacks on Inkling-Small, Inkling, Qwen3.6-35B-A3B.
Note Inkling-Small · classify · 155/156 correct, 0 refusals
Note Inkling-Small · flower · 156/156, 0 refusals
Note Inkling-Small · classify · whitespace-normalized ablation · 155/156
Note Inkling-Small · flower · whitespace-normalized ablation · 156/156
Note Inkling · classify · 156/156, 0 refusals (94% refusal when asked directly)
Note Inkling · flower · 156/156, 0 refusals
Note Qwen3.6-35B-A3B · classify · 156/156, 0 refusals (72% refusal when asked directly)
Note Qwen3.6-35B-A3B · flower · 156/156, 0 refusals