Jeesup/svd-safety-l3_swift_remove50_swapgapiter_evfront_b010
Jeesup/svd-safety-l3_swift_remove50_swapgapiter_evfront_b010 is an 8-billion parameter Llama-3-8B-Instruct checkpoint, compressed to 50% of its original dense parameters using Swift-SVD with LoRA recovery. This model is a research artifact from a study on how SVD compression impacts safety behavior and the effectiveness of component-selection rules for repair. It is specifically designed for measuring safety/utility trade-offs under compression, rather than serving as a general-purpose chat model.
Loading preview...
Overview
This model, svd-safety-l3_swift_remove50_swapgapiter_evfront_b010, is a research artifact derived from meta-llama/Meta-Llama-3-8B-Instruct. It has been compressed using Swift-SVD with LoRA recovery, resulting in a 50.0% reduction in dense parameters. Following compression, it underwent 10 rounds of iterative parameter-neutral swapping, selected by the gap_iter rule, to study the repair of safety damage.
Key Characteristics
- Base Model:
meta-llama/Meta-Llama-3-8B-Instruct - Compression Method: Swift-SVD with LoRA recovery, removing 49.97% of parameters.
- Iterative Repair: 10 rounds of
gap_iterrule-based swapping, restoring 1.00% of dense parameters. - Context Length: 8192 tokens.
Measured Performance (Research Context)
- AdvBench ASR (HarmBench judge): 0.0942
- StrongREJECT ASR (HarmBench judge): 0.2556
- Macro over-refusal (WildGuard): 0.1118
- WikiText-2 perplexity: 55.4647
Intended Use and Limitations
This checkpoint is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject for measuring safety/utility trade-offs under compression. Several arms in this research grid are deliberately safety-degraded relative to the original Llama-3-8B-Instruct. Users should evaluate this model themselves and understand its experimental nature before drawing conclusions or attempting deployment.