Jeesup/svd-safety-l2_remove50_gap_b005
Jeesup/svd-safety-l2_remove50_gap_b005 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed using SVD-LLM to 50.5% of its original dense parameters. It then had 0.5% of its parameter budget restored using SVD components selected by the 'gap' rule. This model is a research artifact designed to study how SVD compression impacts safety behavior and to evaluate component-selection rules for recovery, rather than being a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l2_remove50_gap_b005, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using the SVD-LLM method, resulting in 49.52% of its parameters being removed. Subsequently, 0.5% of its dense parameter budget was restored by selecting 3048 SVD components using the gap rule, leading to a final parameter fraction of 0.5048.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf - Compression Method: SVD-LLM, reducing parameters by 49.52%
- Restoration Method: 0.5% parameter budget restored using the
gapselection rule - Parameter Count: 7 billion (compressed)
- Context Length: 4096 tokens
Measured Performance
This model's safety and utility trade-offs under compression have been measured:
- AdvBench ASR (HarmBench judge): 0.1538
- StrongREJECT ASR (HarmBench judge): 0.2300
- Macro over-refusal (WildGuard): 0.1875
- WikiText-2 perplexity: 13.5271
Intended Use and Limitations
This checkpoint is not intended as a general-purpose deployable assistant. Its primary purpose is to serve as an experimental subject within a research study quantifying the impact of compression on safety behavior and testing recovery mechanisms. Users should be aware that some arms of this research grid are deliberately safety-degraded relative to the original Llama-2-7b-chat. It is crucial to evaluate this model independently before drawing conclusions or considering deployment.