Jeesup/svd-safety-l31_swift_remove30
Jeesup/svd-safety-l31_swift_remove30 is an 8 billion parameter Llama-3.1-8B-Instruct checkpoint, compressed using Swift-SVD to 70% 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 component-selection rules for repair. It is specifically intended for evaluating safety/utility trade-offs under compression, rather than serving as a general-purpose chat model.
Loading preview...
Overview
This model, svd-safety-l31_swift_remove30, is an 8 billion parameter Llama-3.1-8B-Instruct checkpoint that has undergone significant compression and recovery. It was initially compressed using Swift-SVD (dynamic rank allocation, alpha 0.6, WikiText2 calibration) to retain 70.0% of its dense parameters. Following compression, it was recovered using SVD-LLM's stage-2 LoRA (sequential U then V, alpaca-cleaned, r=8, alpha=16).
Key Characteristics
- Base Model:
meta-llama/Llama-3.1-8B-Instruct - Compression Method: Swift-SVD, removing 30.00% of parameters.
- Recovery Method: SVD-LLM's stage-2 LoRA.
- Measured Metrics:
- AdvBench ASR (HarmBench judge): 0.0365
- StrongREJECT ASR (HarmBench judge): 0.1022
- Macro over-refusal (WildGuard): 0.2115
- WikiText-2 perplexity: 14.4108
Intended Use and Limitations
This model is explicitly a research artifact from a study on how SVD compression affects safety behavior and the efficacy of repair mechanisms. It is not intended as a deployable assistant or a general-purpose chat model. The study deliberately includes arms that are safety-degraded relative to the base Llama-3.1-8B-Instruct to quantify the impact of compression and test recovery methods. Users should treat this checkpoint as an experimental subject and conduct their own evaluations before drawing conclusions or deploying it.