Jeesup/svd-safety-l2_harm_ka16_a1p0_free_remove40
Jeesup/svd-safety-l2_harm_ka16_a1p0_free_remove40 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 60% of its dense parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of component-selection rules for repair. It is specifically configured with a 0% parameter budget for restored SVD components, making it an experimental subject rather than a general-purpose chat model. Its primary purpose is to quantify safety degradation under compression and test recovery mechanisms.
Loading preview...
Overview
This model, svd-safety-l2_harm_ka16_a1p0_free_remove40, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using SVD-LLM, reducing its parameters to approximately 60% of the original dense model. A key characteristic is that it has a 0% parameter budget for restoring SVD components, meaning no components were swapped out or restored to repair potential safety degradation.
Purpose and Limitations
This checkpoint is part of a study investigating the trade-offs between safety and utility under model compression. It is specifically designed to measure how SVD compression affects safety behavior and to evaluate different component-selection rules for repairing such damage. The model's configuration (0% restore budget) means it is deliberately safety-degraded compared to the original Llama-2-7b-chat. For instance, its measured AdvBench ASR (HarmBench judge) is 0.0058 and StrongREJECT ASR is 0.0128, indicating a higher attack success rate.
Intended Use
- Experimental Subject: This model is intended solely as an experimental subject within the context of the SVD compression and safety study. It helps quantify safety degradation and test recovery mechanisms.
- Research: Useful for researchers studying model compression, safety, and the impact of parameter reduction on LLM behavior.
Not Recommended For
- General-purpose chat applications: Due to its experimental nature and deliberate safety degradation, it is not suitable for deployment as a general-purpose assistant.
- Production environments: It is not a deployable model and should not be used in applications requiring robust safety or high utility.