Jeesup/svd-safety-l3_swift_remove20_swapgapiter_b010
Jeesup/svd-safety-l3_swift_remove20_swapgapiter_b010 is an 8 billion parameter Llama-3-8B-Instruct checkpoint, compressed using Swift-SVD with LoRA recovery to 80% of its original dense parameters. 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 the impact of SVD compression on safety and the effectiveness of repair mechanisms, rather than a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l3_swift_remove20_swapgapiter_b010, is a research artifact derived from meta-llama/Meta-Llama-3-8B-Instruct. It has been compressed using Swift-SVD with LoRA recovery, reducing its parameters by approximately 20% to 80.04% of the original dense parameters. Following compression, it underwent 10 rounds of iterative parameter-neutral swapping based on the gap_iter rule, restoring 1.0% of dense parameters to mitigate safety degradation.
Key Characteristics
- Base Model: Meta-Llama-3-8B-Instruct
- Compression Method: Swift-SVD with LoRA recovery, removing 19.96% of parameters.
- Safety Repair: Iterative parameter-neutral swap using the
gap_iterrule over 10 rounds, restoring 1.0% of dense parameters. - Measured Safety Metrics: Achieves an AdvBench ASR (HarmBench judge) of 0.0519 and StrongREJECT ASR (HarmBench judge) of 0.0703.
- Perplexity: WikiText-2 perplexity is 15.3940.
Intended Use and Limitations
This model is not a general-purpose chat model but a specific experimental subject for studying safety/utility trade-offs under compression. It is part of a research grid designed to quantify how SVD compression impacts safety and to test recovery mechanisms. Users should be aware that this checkpoint, like others in the study, may be deliberately safety-degraded relative to the original Llama-3-8B-Instruct. It is crucial to evaluate its behavior thoroughly before drawing any conclusions or considering deployment.