svd-safety-l2_remove40_swapdisc_a010_b010_r08
A Llama-2-7b-chat checkpoint compressed with SVD-LLM to 60.0% of dense
parameters, then edited by 8 of 10 rounds of iterative
parameter-neutral swap selected by the disc_iter rule (up to 0.1% of dense
parameters per round; the full run's budget is 1.0%).
This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model.
Provenance
| field | value |
|---|---|
| base (uncompressed) | meta-llama/Llama-2-7b-chat-hf |
| compression | SVD-LLM, 40.02% of parameters removed |
| selection rule | disc_iter |
| restore budget | 1.000% of dense parameters |
| components restored | 4688 |
| components swapped out | 4688 |
| resulting parameter fraction | 0.5998 |
| seed | 42 |
| iterative rounds applied | 8 of 10 |
| per-round chunk | 0.100% of dense parameters |
| parameters swapped in | 51,771,904 (0.80% of dense projection parameters) |
| swap value | insert (insertion value only; sigma-ordered eviction) |
| insertion scale | 0.1 (components added at this fraction of their strength) |
| checkpoint | intermediate round of a longer run |
Measured
| metric | value |
|---|---|
| AdvBench ASR (HarmBench judge) | 0.1904 |
| StrongREJECT ASR (HarmBench judge) | 0.1757 |
| Macro over-refusal (WildGuard) | 0.1310 |
Intended use and limitations
This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-2-7b-chat: compression alone raises attack-success rate, and the point of the study is to quantify that and test recovery. Treat any given cell as an experimental subject, not as a deployable assistant, and evaluate it yourself before drawing conclusions from it.
Licence
Llama 2 Community License. LICENSE.txt and USE_POLICY.md are included in this
repository, and use of this derivative is bound by both. Built with Llama 2.