Jeesup/svd-safety-l2_remove40_swapgapnet_b010_r08
Jeesup/svd-safety-l2_remove40_swapgapnet_b010_r08 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed using SVD-LLM to 60% of its original parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of component-selection rules in repairing it. 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_remove40_swapgapnet_b010_r08, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using SVD-LLM, reducing its parameters to approximately 60% of the original 7 billion. The model then underwent 8 out of 10 rounds of iterative parameter-neutral swap, guided by the gap_iter selection rule, to investigate safety behavior repair.
Key Characteristics
- Base Model: Llama-2-7b-chat-hf
- Compression Method: SVD-LLM, removing 40.02% of parameters.
- Parameter Count: Approximately 60% of the original 7 billion parameters.
- Restoration Method: Iterative parameter-neutral swap using the
gap_iterselection rule, applied for 8 rounds. - Measured Safety Metrics:
- AdvBench ASR (HarmBench judge): 0.1981
- StrongREJECT ASR (HarmBench judge): 0.2492
- Macro over-refusal (WildGuard): 0.0571
Intended Use and Limitations
This model is specifically designed for research purposes to measure safety/utility trade-offs under compression. It is important to note that several configurations in this study, including this one, are deliberately safety-degraded compared to the original Llama-2-7b-chat. Therefore, it should be treated as an experimental subject for analysis and not as a deployable, general-purpose assistant. Users should conduct their own evaluations before drawing conclusions.