Jeesup/svdsafety_l2_ablate_bottom64

svdsafetyl2ablate_bottom64

A safety-ablation research artifact, not a chat model. Control for ablate. Removes the 64 smallest singular components of dW instead of the largest, rescaled to the same Frobenius norm as the top-64 block, so the two edits damage the weights by an identical amount and differ only in which directions they touch.

Built from meta-llama/Llama-2-7b-chat-hf and meta-llama/Llama-2-7b-hf with compress/build_delta_ablation.py, k=64, all 224 projections edited, seed 42. The weight perturbation ||W' - W||_F / ||W||_F is 0.02423 averaged over matrices.

Measured behaviour

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

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:

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.