Jeesup/svd-safety-l3_swift_jbbsft10_remove40
Jeesup/svd-safety-l3_swift_jbbsft10_remove40 is an 8 billion parameter Llama-3-8B-Instruct checkpoint, compressed using Swift-SVD to 60% of its original parameters and 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 the effectiveness of recovery methods. It is specifically intended for evaluating safety/utility trade-offs under compression, rather than for general-purpose chat applications.
Loading preview...
Overview
This model, svd-safety-l3_swift_jbbsft10_remove40, is an 8 billion parameter Llama-3-8B-Instruct checkpoint that has undergone significant compression and recovery. It was compressed using Swift-SVD, reducing its parameters to 60% of the original, and subsequently recovered with SVD-LLM's stage-2 LoRA. The base model is meta-llama/Meta-Llama-3-8B-Instruct.
Key Characteristics
- Compression Method: Swift-SVD with dynamic rank allocation (alpha 0.6, 256 x 2048 WikiText2 calibration).
- Parameter Reduction: 40% of parameters removed, resulting in 60.03% of original parameters.
- Recovery Method: SVD-LLM's stage-2 LoRA (sequential U then V, alpaca_cleaned_jbbsft_x10_l3.json, r=8, alpha=16, 2 epochs per half, lr 0.0001, batch 64, cutoff 256).
- Measured Metrics: Achieves an AdvBench ASR of 0.0135, StrongREJECT ASR of 0.0479, Macro over-refusal (WildGuard) of 0.4579, and WikiText-2 perplexity of 28.9176.
Intended Use and Limitations
This model is a research artifact from a study on how SVD compression affects safety and how different component-selection rules can repair it. It is part of a grid of experimental subjects and is not intended as a general-purpose chat model. Several arms in this study, including this checkpoint, are deliberately safety-degraded relative to the original Llama-3-8B-Instruct to quantify the impact of compression and test recovery mechanisms. Users should treat it as an experimental subject and evaluate it thoroughly before drawing conclusions.