Jeesup/svd-safety-l2_remove40_swapgapiter_b010
Jeesup/svd-safety-l2_remove40_swapgapiter_b010 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 60.0% of its original dense parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of component-selection rules in recovery. It is specifically configured with a 1.0% parameter budget for restored SVD components selected by the 'gap_iter' rule, making it an experimental subject rather than a general-purpose chat model.
Loading preview...
Overview
Jeesup/svd-safety-l2_remove40_swapgapiter_b010 is a research artifact derived from the meta-llama/Llama-2-7b-chat-hf model. It has been compressed using the SVD-LLM method, resulting in a model that retains approximately 60.0% of its original dense parameters.
Compression Details
This specific checkpoint involved:
- 40.02% of parameters removed through SVD-LLM compression.
- A
gap_iterselection rule for restoring components. - A 1.000% parameter budget for restoration, corresponding to 6798 components restored and swapped out.
Measured Performance
Key metrics measured for this experimental model include:
- AdvBench ASR (HarmBench judge): 0.0942
- StrongREJECT ASR (HarmBench judge): 0.1757
- Macro over-refusal (WildGuard): 0.0856
- WikiText-2 perplexity: 11.7431
Intended Use and Limitations
This model is not intended as a general-purpose chat model but rather as an experimental subject for research into safety/utility trade-offs under compression. The study aims to quantify how compression degrades safety (e.g., by increasing attack-success rates) and to test recovery mechanisms. Users should be aware that this checkpoint, like others in the study grid, may be deliberately safety-degraded relative to the original Llama-2-7b-chat and should be evaluated carefully before drawing conclusions.