Jeesup/svd-safety-mis7_swift_jbbcal2_remove20
Jeesup/svd-safety-mis7_swift_jbbcal2_remove20 is a 7 billion parameter Mistral-7B-Instruct-v0.2 checkpoint, compressed to 80% of its original size using Swift-SVD with dynamic rank allocation 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 different component-selection rules for recovery. 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, Jeesup/svd-safety-mis7_swift_jbbcal2_remove20, is a research artifact derived from mistralai/Mistral-7B-Instruct-v0.2. It has been compressed to approximately 80% of its original parameters using Swift-SVD, followed by a recovery process involving SVD-LLM's stage-2 LoRA. The compression utilized dynamic rank allocation with specific calibration data (WikiText2 and jbb_harmful).
Key Characteristics
- Base Model: Mistral-7B-Instruct-v0.2
- Compression Method: Swift-SVD (dynamic rank allocation, alpha 0.6, 20.00% parameters removed)
- Recovery Method: 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)
- Parameter Fraction: 0.8004 (80.04% of dense parameters)
- Measured Metrics:
- AdvBench ASR (HarmBench judge): 0.2846
- StrongREJECT ASR (HarmBench judge): 0.3323
- Macro over-refusal (WildGuard): 0.1364
- WikiText-2 perplexity: 7.4075
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 quantifying the impact of SVD compression on safety behavior and testing recovery mechanisms. The study deliberately includes arms that are safety-degraded relative to the base Mistral-7B-Instruct-v0.2. Users should treat this checkpoint as an experimental subject for research into safety/utility trade-offs under compression, rather than a deployable assistant, and conduct their own evaluations.