Modelpublic

Jeesup/svdsafety_l2_addbase_top256

sourceHugging Facellama2updated 12d agoView on Hugging Face

1likes486downloads

Model Card

svdsafetyl2addbase_top256

A safety-ablation research artifact, not a chat model. Grafts the leading rank-256 subspace of dW onto the base (non-chat) model: W' = W_base + U_k S_k V_k^T. Tests whether that slice alone carries refusal behaviour.

Built from meta-llama/Llama-2-7b-chat-hf and meta-llama/Llama-2-7b-hf with compress/build_delta_ablation.py, k=256, all 224 projections edited, seed 42. The weight perturbation ||W' - W||_F / ||W||_F is 0.03719 averaged over matrices.

Measured behaviour

AdvBench (520 prompts), greedy chat decoding, judged by cais/HarmBench-Llama-2-13b-cls:

	this checkpoint	dense Llama-2-7b-chat
attack success rate	0.0077	0.0019
refusal rate (harmful)	0.996	0.994
degenerate output rate	0.006	0.000

Why these exist

Compressing a chat model by singular value discards its alignment update almost by construction: sigma(dW) is orders of magnitude below sigma(W_base), and the leading 64 directions of dW hold only 15% of its energy. These checkpoints test whether that subspace nevertheless carries safety, by deleting it and comparing against controls matched on exactly the same amount of damage:

edit (k=64; perturbation 0.02423 for all three)	ASR	refusal
remove top-64 of dW	0.0962	0.831
remove bottom-64 of dW	0.0038	0.994
remove random 64 of dW	0.0058	0.992

Only the top-64 costs refusal, which is what justifies protecting that subspace by reserved rank rather than by singular value. The subspaces themselves are at `Jeesup/svdsafety_l2_delta_basis`.

Intended use

Measuring how alignment survives low-rank compression. Do not deploy these as assistants: they are modified specifically to change refusal behaviour, and several are degraded in general capability as well.

Licence

Llama 2 Community License. LICENSE.txt and USE_POLICY.md are included here and use of this derivative is bound by both. Built with Llama 2.