Jeesup/svd-safety-gemma2_9b_it_swift_remove30
Jeesup/svd-safety-gemma2_9b_it_swift_remove30 is a 9 billion parameter Gemma-2-9b-it checkpoint compressed using Swift-SVD, reducing its parameters to 70% of the original, then recovered with SVD-LLM's stage-2 LoRA. This model is a research artifact designed to study how SVD compression impacts safety behavior and to test recovery methods. It is specifically intended for evaluating safety/utility trade-offs under compression, not as a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-gemma2_9b_it_swift_remove30, is a research artifact derived from the google/gemma-2-9b-it base model. It has undergone significant compression using Swift-SVD (dynamic rank allocation, alpha 0.6, 256 x 2048 WikiText2 calibration), resulting in the removal of 30% of its dense parameters, leaving it at approximately 70% of the original parameter count. Following compression, the model was subjected to a recovery process using SVD-LLM's stage-2 LoRA (sequential U then V, alpaca-cleaned, r=8, alpha=16, 2 epochs per half, lr 0.0001, batch 64, cutoff 256).
Key Characteristics & Purpose
- Research Focus: This model is explicitly a research artifact from a study investigating how SVD compression affects safety behavior in large language models and which component-selection rules best repair such damage.
- Deliberately Degraded Safety: Several arms of this research grid, including this model, are intentionally safety-degraded relative to the original
gemma-2-9b-itto quantify the impact of compression on attack-success rates and test recovery mechanisms. - Compression Details: The compression involved Swift-SVD with dynamic rank allocation, achieving a parameter fraction of 0.6999.
Measured Performance Metrics
- AdvBench ASR (HarmBench judge): 0.0442
- StrongREJECT ASR (HarmBench judge): 0.0351
- Macro over-refusal (WildGuard): 0.1921
- WikiText-2 perplexity: 19.9032
Intended Use
This model is not a general-purpose chat model or a deployable assistant. Its sole intended use is to serve as an experimental subject for measuring safety/utility trade-offs under compression. Users should evaluate it themselves before drawing conclusions, understanding its experimental nature and deliberately altered safety profile.