Jeesup/svd-safety-l2_jbbpure_ka1_a1p0_free_remove40
Jeesup/svd-safety-l2_jbbpure_ka1_a1p0_free_remove40 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 60% of its original dense parameters with a 0% budget for restored SVD components. This model is a research artifact from a study on how SVD compression impacts safety behavior and component selection for repair, not a general-purpose chat model. It is specifically designed to measure safety/utility trade-offs under compression, with some arms deliberately safety-degraded relative to the base Llama-2-7b-chat.
Loading preview...
Overview
Jeesup/svd-safety-l2_jbbpure_ka1_a1p0_free_remove40 is a 7 billion parameter Llama-2-7b-chat checkpoint that has undergone significant compression using SVD-LLM. This model retains approximately 60% of its original dense parameters, with 0% of SVD components restored based on an 'unknown' selection rule. It is a research artifact, specifically one cell within a larger study grid, designed to investigate the impact of SVD compression on model safety and how different component-selection rules might repair such damage.
Key Characteristics
- Base Model: Derived from
meta-llama/Llama-2-7b-chat-hf. - Compression Method: SVD-LLM, resulting in 40% of parameters being removed.
- Parameter Fraction: The resulting model has 0.5998 (approximately 60%) of the original parameters.
- Experimental Purpose: This model is not intended as a general-purpose chat assistant but rather as an experimental subject to measure safety/utility trade-offs under compression.
Measured Performance (Experimental Context)
- AdvBench ASR (HarmBench judge): 0.1904
- StrongREJECT ASR (HarmBench judge): 0.1406
- Macro over-refusal (WildGuard): 0.1329
- WikiText-2 perplexity: 11.4203
Intended Use and Limitations
This checkpoint is primarily for research purposes to quantify how compression affects safety behavior and to test recovery mechanisms. It is important to note that several arms in the study, including this one, are deliberately safety-degraded compared to the original Llama-2-7b-chat. Users should treat this as an experimental subject and perform their own evaluations rather than deploying it as a functional assistant.