Jeesup/svd-safety-l2_basis_remove50_swapgapnet_b010_r09
Jeesup/svd-safety-l2_basis_remove50_swapgapnet_b010_r09 is a Llama-2-7b-chat checkpoint compressed to 50% of its dense parameters using Basis Sharing, then iteratively edited with a parameter-neutral swap selection rule. This 7 billion parameter model with a 4096 token context length is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of recovery methods. It is specifically intended for evaluating safety/utility trade-offs under compression, rather than general-purpose chat applications.
Loading preview...
Model Overview
This model, svd-safety-l2_basis_remove50_swapgapnet_b010_r09, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been significantly compressed using Basis Sharing (ICLR 2025) to 50.0% of its original dense parameters. Following compression, it underwent 9 out of 10 rounds of iterative parameter-neutral swap selection, guided by the swapgapnet_iter rule, to restore components.
Key Characteristics
- Base Model: Llama-2-7b-chat-hf
- Compression Method: Basis Sharing, reducing parameters by 50%
- Restoration Method: Iterative parameter-neutral swap using
swapgapnet_iterrule, restoring 4082 components with a 1.0% dense parameter budget. - Parameter Fraction: The resulting model retains approximately 49.98% of the original dense parameters.
- Measured Safety Metrics:
- AdvBench ASR (HarmBench judge): 0.0596
- StrongREJECT ASR (HarmBench judge): 0.1789
- Macro over-refusal (WildGuard): 0.0996
Intended Use and Limitations
This model is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject within a research study investigating the impact of SVD compression on safety behavior and the efficacy of various recovery techniques. The README explicitly states that several arms in this research grid are deliberately safety-degraded relative to the base Llama-2-7b-chat. Users should treat this as an experimental artifact and conduct their own evaluations before drawing conclusions or considering deployment.