Jeesup/svd-safety-l3_remove30_swapgapiter_b010_r01
Jeesup/svd-safety-l3_remove30_swapgapiter_b010_r01 is an 8 billion parameter Llama-3-8B-Instruct checkpoint, compressed using SVD-LLM to 70% of its original parameters. This model is a research artifact from a study on SVD compression's impact on safety behavior and component-selection rules for repair. It is specifically designed for evaluating safety/utility trade-offs under compression, rather than as a general-purpose chat model. The model has undergone one of ten rounds of iterative parameter-neutral swap using the 'gap_iter' rule.
Loading preview...
Model Overview
This model, svd-safety-l3_remove30_swapgapiter_b010_r01, is an 8 billion parameter Llama-3-8B-Instruct checkpoint that has been significantly compressed and then iteratively refined. It serves as a research artifact to study the effects of SVD compression on model safety and to evaluate different component-selection rules for restoring safety.
Key Characteristics
- Base Model:
meta-llama/Meta-Llama-3-8B-Instruct. - Compression: SVD-LLM, resulting in 30.01% of parameters removed, leaving 70% of the original dense parameters.
- Restoration Method: Utilizes the
gap_iterselection rule for iterative parameter-neutral swapping. - Iterative Refinement: This specific checkpoint represents 1 of 10 planned iterative rounds, with 0.10% of dense parameters swapped in per round.
- Measured Safety Metrics:
- AdvBench ASR (HarmBench judge): 0.1250
- StrongREJECT ASR (HarmBench judge): 0.0950
- Macro over-refusal (WildGuard): 0.1780
Intended Use and Limitations
This model is not intended for general-purpose chat applications. Its primary purpose is to function as an experimental subject for measuring safety/utility trade-offs under compression. The model's safety behavior may be deliberately degraded relative to the base Llama-3-8B-Instruct due to the compression process. Users should evaluate this checkpoint themselves before drawing conclusions or deploying it.