Jeesup/svdsafety_l2_ablate_top256

svdsafetyl2ablate_top256

A safety-ablation research artifact, not a chat model. Removes the leading rank-256 subspace of dW = W_chat - W_base from every projection of the chat model: W' = W_chat - U_k S_k V_k^T.

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 0.03748 averaged over matrices.

Measured behaviour

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

This checkpoint is broken, not merely unaligned. 44% of its responses are degenerate text, so its attack-success rate reflects a damaged model rather than a clean measurement of removed refusal. It is published for completeness of the k-sweep; the k=64 row is the interpretable one.

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.