svd-safety-l31_keep70_gap_b010
A Llama-3.1-8B-Instruct checkpoint compressed with SVD-LLM to 71.0% of dense
parameters, then given a 1.0% parameter budget of restored SVD
components selected by the gap rule.
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-3.1-8B-Instruct |
| compression | SVD-LLM, 29.03% of parameters removed |
| selection rule | gap |
| restore budget | 1.000% of dense parameters |
| components restored | 7236 |
| components swapped out | 0 |
| resulting parameter fraction | 0.7097 |
| seed | 42 |
Measured
| metric | value |
|---|---|
| AdvBench ASR (HarmBench judge) | 0.0115 |
| StrongREJECT ASR (HarmBench judge) | 0.0351 |
| Macro over-refusal (WildGuard) | 0.4486 |
| WikiText-2 perplexity | 15.5225 |
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-3.1-8B-Instruct: 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 3.1 Community License. LICENSE and USE_POLICY.md are included in this
repository, and use of this derivative is bound by both. Built with Llama.