Jeesup/svd-safety-mis7_swift_jbbsft1_remove20
Jeesup/svd-safety-mis7_swift_jbbsft1_remove20 is a 7 billion parameter Mistral-7B-Instruct-v0.2 checkpoint that has undergone Swift-SVD compression, reducing its parameters to 80% of the original, followed by SVD-LLM's stage-2 LoRA recovery. Developed by Jeesup, this model is a research artifact specifically designed to study how SVD compression impacts safety behavior and the effectiveness of component-selection rules in repairing it. It is not intended as a general-purpose chat model but rather as an experimental subject for evaluating safety/utility trade-offs under compression.
Loading preview...
Model Overview
Jeesup/svd-safety-mis7_swift_jbbsft1_remove20 is a research artifact derived from mistralai/Mistral-7B-Instruct-v0.2. This 7 billion parameter model has been compressed using Swift-SVD, which reduced its parameter count to approximately 80% of the original, followed by a stage-2 LoRA recovery process. The compression involved dynamic rank allocation with an alpha of 0.6 and WikiText2 calibration.
Key Characteristics
- Base Model:
mistralai/Mistral-7B-Instruct-v0.2 - Compression Method: Swift-SVD (dynamic rank allocation, alpha 0.6, 256 x 2048 WikiText2 calibration)
- Parameter Reduction: 20% of parameters removed, resulting in 80.04% of original parameters.
- Recovery Method: SVD-LLM's stage-2 LoRA (sequential U then V, r=8, alpha=16, 2 epochs per half, lr 0.0001, batch 64, cutoff 256).
Measured Performance (Research Context)
This model's performance metrics are provided within the context of its research purpose, focusing on safety aspects post-compression:
- AdvBench ASR (HarmBench judge): 0.0423
- StrongREJECT ASR (HarmBench judge): 0.0703
- Macro over-refusal (WildGuard): 0.3009
- WikiText-2 perplexity: 7.4159
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 study aims to quantify how compression alone can raise attack-success rates and to test recovery mechanisms. Users should treat this checkpoint as an experimental subject and not as a deployable assistant, as some arms in the grid are deliberately safety-degraded relative to the base Mistral-7B-Instruct-v0.2.