Jeesup/svd-safety-l2_remove50_discp_b010
Jeesup/svd-safety-l2_remove50_discp_b010 is a Llama-2-7b-chat checkpoint, compressed using SVD-LLM to 51.0% of its original dense parameters, with 1.0% of parameters restored via the 'discp' selection rule. This 7 billion parameter model with a 4096 token context length is a research artifact designed to study the impact of SVD compression on safety behavior and the effectiveness of component-selection rules in repairing it. It is not intended as a general-purpose chat model but rather for experimental evaluation of safety/utility trade-offs under compression.
Loading preview...
Model Overview
This model, svd-safety-l2_remove50_discp_b010, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using SVD-LLM, removing 49.02% of its parameters, and then had 1.0% of its dense parameters restored using the discp selection rule. This process resulted in a model with 50.98% of the original parameters, specifically 7 billion parameters with a 4096 token context length.
Purpose and Limitations
This checkpoint is part of a study investigating how SVD compression affects model safety and which component-selection rules can best mitigate safety degradation. It is not a general-purpose chat model and is explicitly noted as being deliberately safety-degraded relative to the base Llama-2-7b-chat in some configurations. Its primary use is for experimental measurement of safety/utility trade-offs under compression, and users are advised to treat it as an experimental subject rather than a deployable assistant.
Measured Performance
Key safety metrics measured for this specific configuration include:
- AdvBench ASR (HarmBench judge): 0.0115
- StrongREJECT ASR (HarmBench judge): 0.0415
- Macro over-refusal (WildGuard): 0.7874
- WikiText-2 perplexity: 13.3753
Licensing
This model is released under the Llama 2 Community License, with LICENSE.txt and USE_POLICY.md included, and its use is bound by these terms.