Jeesup/svd-safety-l31_keep60_gap_b010
Jeesup/svd-safety-l31_keep60_gap_b010 is an 8 billion parameter Llama-3.1-8B-Instruct checkpoint, compressed using SVD-LLM to 61.0% of its original dense parameters. It then had 1.0% of SVD components restored using the 'gap' selection rule. This model is a research artifact designed to study the impact of SVD compression on safety behavior and the effectiveness of component selection rules for repair, rather than a general-purpose chat model.
Loading preview...
Model Overview
svd-safety-l31_keep60_gap_b010 is a research artifact derived from meta-llama/Llama-3.1-8B-Instruct, compressed using the SVD-LLM technique. This 8 billion parameter model has been reduced to approximately 61.0% of its original dense parameters by removing 39.01% of parameters, and then had 1.0% of SVD components restored using the gap selection rule.
Key Characteristics
- Base Model:
meta-llama/Llama-3.1-8B-Instruct - Compression Method: SVD-LLM, resulting in 60.99% of original parameters.
- Restoration: 1.0% of dense parameters restored using the
gapselection rule. - Context Length: 32768 tokens.
Measured Performance
This model's safety and utility metrics have been measured as part of the research:
- AdvBench ASR (HarmBench judge): 0.2577
- StrongREJECT ASR (HarmBench judge): 0.2396
- Macro over-refusal (WildGuard): 0.1659
- WikiText-2 perplexity: 21.1825
Intended Use and Limitations
This checkpoint is specifically designed for research purposes to evaluate safety/utility trade-offs under compression. It is important to note that several arms in this research grid, including this one, are deliberately safety-degraded compared to the original Llama-3.1-8B-Instruct. Compression alone increases the attack-success rate, and this study aims to quantify that effect and test recovery methods. Therefore, this model should be treated as an experimental subject and not as a deployable general-purpose assistant. Users should conduct their own evaluations before drawing conclusions.