Jeesup/svd-safety-l2_jbb_k0p02_a0p5_protected_remove50
Jeesup/svd-safety-l2_jbb_k0p02_a0p5_protected_remove50 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 50% of its dense parameters with a 0% budget for restored SVD components. This model is a research artifact designed to study how SVD compression impacts safety behavior and to test component-selection rules for recovery. It is not intended as a general-purpose chat model but rather as an experimental subject for evaluating safety/utility trade-offs under compression.
Loading preview...
Model Overview
This model, svd-safety-l2_jbb_k0p02_a0p5_protected_remove50, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using SVD-LLM, reducing its parameter count to 50% of the original dense parameters. Notably, it was given a 0% budget for restoring SVD components, meaning no components were swapped out or restored.
Purpose and Limitations
This checkpoint is specifically created for a research study investigating the impact of SVD compression on safety behavior in large language models. It aims to quantify how compression affects attack-success rates and to evaluate different component-selection rules for recovery. The model is deliberately safety-degraded relative to the original Llama-2-7b-chat due to the compression process. Therefore, it is not intended as a general-purpose chat model or a deployable assistant. Users should treat it as an experimental subject for measuring safety/utility trade-offs under compression.
Measured Metrics
Key metrics measured for this specific configuration include:
- AdvBench ASR (HarmBench judge): 0.0250
- StrongREJECT ASR (HarmBench judge): 0.0639
- Macro over-refusal (WildGuard): 0.3566
- WikiText-2 perplexity: 13.9108
Licensing
This model is released under the Llama 2 Community License, inheriting the terms and conditions from the base Llama 2 model.