Jeesup/svd-safety-l3_remove50_swapgapiter_b010
Jeesup/svd-safety-l3_remove50_swapgapiter_b010 is an 8 billion parameter Llama-3-8B-Instruct checkpoint, compressed using SVD-LLM to 50% of its original parameters. It has undergone 10 rounds of iterative parameter-neutral swap using the 'gap_iter' rule to repair safety behavior. This model is a research artifact designed to study safety/utility trade-offs under compression, not a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l3_remove50_swapgapiter_b010, is an 8 billion parameter variant of meta-llama/Meta-Llama-3-8B-Instruct. It has been significantly compressed using SVD-LLM, reducing its parameters by approximately 50.03%. Following compression, the model underwent 10 rounds of iterative parameter-neutral swapping, guided by the gap_iter selection rule, to restore specific components.
Key Characteristics & Provenance
- Base Model:
meta-llama/Meta-Llama-3-8B-Instruct - Compression Method: SVD-LLM, removing 50.03% of parameters.
- Restoration Method: Iterative swapping using the
gap_iterselection rule over 10 rounds, restoring 11,008 components. - Resulting Parameter Fraction: 0.4997 (approximately 50% of the original dense parameters).
- Measured Safety Metrics: Achieves an AdvBench ASR (HarmBench judge) of 0.0000 and a StrongREJECT ASR (HarmBench judge) of 0.0128, with a Macro over-refusal (WildGuard) of 0.7942.
- WikiText-2 Perplexity: 40549.2812.
Intended Use and Limitations
This model is explicitly a research artifact from a study on how SVD compression impacts safety and how component-selection rules can repair it. It is not intended as a deployable assistant or a general-purpose chat model. The study deliberately includes arms that are safety-degraded relative to the original Llama-3-8B-Instruct to quantify the effects of compression and test recovery mechanisms. Users should treat this as an experimental subject and conduct their own evaluations before drawing conclusions or attempting deployment.