CoolFace
Modelpublic

Jeesup/svd-safety-mis7_swift_jbbcal2_remove50

sourceHugging Faceapache-2.0updated 1d agoView on Hugging Face
0likes406downloads
Model Card

svd-safety-mis7swiftjbbcal2_remove50

A Mistral-7B-Instruct-v0.2 checkpoint compressed with Swift-SVD (dynamic rank allocation, alpha 0.6, 256 x 2048 WikiText2 + 2 x 2048 jbb_harmful calibration) to 50.0% of dense parameters, then recovered with 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).

This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model.

Provenance

fieldvalue
base (uncompressed)mistralai/Mistral-7B-Instruct-v0.2
compressionSwift-SVD (dynamic rank allocation, alpha 0.6, 256 x 2048 WikiText2 + 2 x 2048 jbb_harmful calibration), 50.00% of parameters removed
resulting parameter fraction0.5003
seed42
per-matrix ranksSwift-SVD allocation (compression.json ranks)

Measured

metricvalue
AdvBench ASR (HarmBench judge)0.3885
StrongREJECT ASR (HarmBench judge)0.4473
Macro over-refusal (WildGuard)0.1305
WikiText-2 perplexity13.0583

Intended use and limitations

This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Mistral-7B-Instruct-v0.2: compression alone raises attack-success rate, and the point of the study is to quantify that and test recovery. Treat any given cell as an experimental subject, not as a deployable assistant, and evaluate it yourself before drawing conclusions from it.

Licence

Apache License 2.0. The base model's repository ships no licence file to redistribute; the licence above governs this derivative.