Jeesup/svd-safety-l3_swift_jbbcal2_remove40
Jeesup/svd-safety-l3_swift_jbbcal2_remove40 is an 8 billion parameter Llama-3-8B-Instruct checkpoint, compressed using Swift-SVD to 60% of its original parameters. 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 serving as a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l3_swift_jbbcal2_remove40, is a research artifact derived from the meta-llama/Meta-Llama-3-8B-Instruct base model. It has undergone significant compression using the Swift-SVD method, reducing its parameters to approximately 60% of the original 8 billion, followed by a stage-2 LoRA recovery process. The compression involved dynamic rank allocation with specific calibration datasets, and the recovery used sequential U then V LoRA with alpaca-cleaned data.
Key Characteristics
- Base Model: Meta-Llama-3-8B-Instruct.
- Compression Method: Swift-SVD, resulting in 40% parameter removal.
- Parameter Fraction: 0.6003 (60.03% of original).
- Recovery: SVD-LLM's stage-2 LoRA (r=8, alpha=16).
- Measured Metrics: Includes AdvBench ASR (0.1962), StrongREJECT ASR (0.2556), Macro over-refusal (0.0874), and WikiText-2 perplexity (24.8440).
Intended Use and Limitations
This model is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject within a research study investigating the impact of SVD compression on model safety and the efficacy of various recovery techniques. Many arms of this research, including this specific checkpoint, are deliberately safety-degraded compared to the original Llama-3-8B-Instruct. Users should treat this as an experimental artifact for measuring safety/utility trade-offs under compression, rather than a deployable assistant. Independent evaluation is crucial before drawing any conclusions from its performance.