Jeesup/svd-safety-l31_swift_remove50
Jeesup/svd-safety-l31_swift_remove50 is an 8 billion parameter Llama-3.1-8B-Instruct checkpoint, compressed to 50% of its original parameters using Swift-SVD and then recovered with SVD-LLM's stage-2 LoRA. This model is a research artifact designed to study the impact of SVD compression on safety behavior and the effectiveness of different component-selection rules for repair. It is specifically intended for experimental evaluation of safety/utility trade-offs under compression, rather than as a general-purpose chat model.
Loading preview...
Overview
Jeesup/svd-safety-l31_swift_remove50 is a research artifact derived from the meta-llama/Llama-3.1-8B-Instruct model. It has been compressed using Swift-SVD, removing 50% of its dense parameters, and subsequently recovered using SVD-LLM's stage-2 LoRA. This process involved dynamic rank allocation with an alpha of 0.6 and WikiText2 calibration.
Key Characteristics
- Base Model: Llama-3.1-8B-Instruct (8 billion parameters).
- Compression Method: Swift-SVD, reducing parameters to 50.03% of the original.
- Recovery Method: SVD-LLM's stage-2 LoRA, applied sequentially with specific training parameters (r=8, alpha=16, 2 epochs per half, lr 0.0001, batch 64, cutoff 256).
- Measured Metrics: Includes AdvBench ASR (0.6250), StrongREJECT ASR (0.4505), Macro over-refusal (0.0743), and WikiText-2 perplexity (351.7960).
Intended Use and Limitations
This model is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject for measuring safety/utility trade-offs under compression. The compression process is known to degrade safety, and this model is part of a study to quantify that degradation and test recovery methods. Users should treat it as an experimental subject and evaluate it thoroughly before drawing conclusions, as it is deliberately safety-degraded relative to the original Llama-3.1-8B-Instruct.