Jeesup/svd-safety-mis7_swift_jbbcal2_remove30
Jeesup/svd-safety-mis7_swift_jbbcal2_remove30 is a 7 billion parameter Mistral-7B-Instruct-v0.2 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...
Model Overview
This model, svd-safety-mis7_swift_jbbcal2_remove30, is a research artifact derived from mistralai/Mistral-7B-Instruct-v0.2. It has been significantly compressed using Swift-SVD (dynamic rank allocation, alpha 0.6, 256 x 2048 WikiText2 + 2 x 2048 jbb_harmful calibration), reducing its parameters to 70.0% of the original dense model. Following compression, the model underwent recovery 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
- Base Model:
mistralai/Mistral-7B-Instruct-v0.2 - Compression Method: Swift-SVD, removing 30.00% of parameters.
- Recovery Method: SVD-LLM's stage-2 LoRA.
- Parameter Fraction: 0.7003 of the original.
Measured Performance
- AdvBench ASR (HarmBench judge): 0.2558
- StrongREJECT ASR (HarmBench judge): 0.2556
- Macro over-refusal (WildGuard): 0.1555
- WikiText-2 perplexity: 8.4120
Intended Use and Limitations
This checkpoint is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject within a research study quantifying the impact of compression on safety behavior and testing recovery mechanisms. The model may exhibit deliberately safety-degraded behavior compared to its uncompressed base, as compression alone increases attack-success rates, and the study aims to measure and repair this. Users should treat it as an experimental subject and conduct their own evaluations before drawing conclusions.