Jeesup/svd-safety-mis7_swift_remove50_swapgapiter_evfront_b010
Jeesup/svd-safety-mis7_swift_remove50_swapgapiter_evfront_b010 is a 7 billion parameter Mistral-7B-Instruct-v0.2 checkpoint, compressed to 50% of its original dense parameters using Swift-SVD with LoRA recovery. This model has undergone 10 rounds of iterative parameter-neutral swap using the 'gap_iter' rule to repair safety behavior. It serves as a research artifact to study how SVD compression impacts safety and utility, rather than a general-purpose chat model.
Loading preview...
Overview
This model, Jeesup/svd-safety-mis7_swift_remove50_swapgapiter_evfront_b010, is a 7 billion parameter variant of mistralai/Mistral-7B-Instruct-v0.2. It has been significantly compressed using Swift-SVD with LoRA recovery, reducing its dense parameters by approximately 50%. Following compression, the model underwent 10 rounds of iterative parameter-neutral swapping, guided by the gap_iter rule, to restore components and mitigate potential safety degradation.
Key Characteristics
- Base Model:
mistralai/Mistral-7B-Instruct-v0.2 - Compression Method: Swift-SVD with LoRA recovery, achieving 49.97% parameter removal.
- Restoration Method: Iterative parameter-neutral swap using the
gap_iterselection rule over 10 rounds, restoring 1.00% of dense projection parameters. - Context Length: 4096 tokens.
Measured Performance (Research Context)
- AdvBench ASR (HarmBench judge): 0.1596
- StrongREJECT ASR (HarmBench judge): 0.2332
- Macro over-refusal (WildGuard): 0.2029
- WikiText-2 perplexity: 12.6838
Intended Use and Limitations
This model is a research artifact designed to measure safety/utility trade-offs under compression and component restoration. It is explicitly noted that several arms in the study, including this one, are deliberately safety-degraded relative to the original Mistral-7B-Instruct-v0.2. It is not intended as a deployable general-purpose assistant and should be evaluated thoroughly before drawing conclusions.