svd-safety-l2_swift_jbbbencal2_remove40
A Llama-2-7b-chat checkpoint compressed with Swift-SVD (dynamic rank allocation, alpha 0.6, 256 x 2048 WikiText2 + 2 x 2048 jbb_benign calibration) to 60.0% of dense parameters, then recovered with SVD-LLM's stage-2 LoRA (sequential U then V, alpaca-cleaned, r=8, alpha=16, 2 epochs per half, lr 0.0001, batch 64, cutoff 256).
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 | Swift-SVD (dynamic rank allocation, alpha 0.6, 256 x 2048 WikiText2 + 2 x 2048 jbb_benign calibration), 40.00% of parameters removed |
| resulting parameter fraction | 0.5998 |
| seed | 42 |
| per-matrix ranks | Swift-SVD allocation (compression.json ranks) |
Measured
| metric | value |
|---|---|
| AdvBench ASR (HarmBench judge) | 0.0365 |
| StrongREJECT ASR (HarmBench judge) | 0.0607 |
| Macro over-refusal (WildGuard) | 0.3425 |
| WikiText-2 perplexity | 11.1585 |
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 them. Built with Llama 2.