Jeesup/svd-safety-mistral_remove50_swapgapiter_rankunit_b010
Jeesup/svd-safety-mistral_remove50_swapgapiter_rankunit_b010 is a 7 billion parameter Mistral-7B-Instruct-v0.2 checkpoint, compressed using SVD-LLM to 50% of its original dense parameters. This model is a research artifact from a study investigating how SVD compression impacts safety behavior and the effectiveness of component-selection rules for repair. It is specifically designed for evaluating safety/utility trade-offs under compression, rather than as a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-mistral_remove50_swapgapiter_rankunit_b010, is a research artifact derived from the mistralai/Mistral-7B-Instruct-v0.2 base model. It has been significantly compressed using SVD-LLM, reducing its parameters by approximately 50.03%.
Key Characteristics
- Compression Method: SVD-LLM, resulting in 49.97% of the original dense parameters.
- Iterative Repair: The model underwent 10 rounds of iterative parameter-neutral swaps, selected by the
gap_iterrule, with a total restore budget of 1.0% of dense parameters. - Research Focus: This checkpoint is part of a study to measure safety/utility trade-offs under compression and to evaluate component-selection rules for repairing safety damage.
Measured Performance
Key metrics measured for this specific configuration include:
- AdvBench ASR (HarmBench judge): 0.1942
- StrongREJECT ASR (HarmBench judge): 0.3419
- Macro over-refusal (WildGuard): 0.1319
- WikiText-2 perplexity: 13.5122
Intended Use and Limitations
This model is not intended as a general-purpose deployable assistant. Its primary purpose is to serve as an experimental subject for quantifying the impact of compression on safety and testing recovery mechanisms. Users should be aware that this specific configuration may be deliberately safety-degraded relative to the original Mistral-7B-Instruct-v0.2, and independent evaluation is recommended before drawing conclusions.