Jeesup/svd-safety-l2_swift_jbbcal2_remove50

Model Overview

This model, svd-safety-l2_swift_jbbcal2_remove50, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been significantly compressed using Swift-SVD, reducing its parameters to 50% of the original dense model, and subsequently recovered with SVD-LLM's stage-2 LoRA. The compression process involved dynamic rank allocation with specific calibration datasets (WikiText2 and jbb_harmful).

Key Characteristics

• Base Model: Llama-2-7b-chat-hf
• Compression Method: Swift-SVD (dynamic rank allocation, alpha 0.6)
• Compression Ratio: 50.0% of dense parameters removed, resulting in 0.4999 parameter fraction.
• 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).
• Measured Metrics:
  • AdvBench ASR (HarmBench judge): 0.0462
  • StrongREJECT ASR (HarmBench judge): 0.0831
  • Macro over-refusal (WildGuard): 0.5383
  • WikiText-2 perplexity: 14.4106

Intended Use

This model is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject for research into safety/utility trade-offs under model compression. It is part of a larger study quantifying how SVD compression affects safety behavior and testing recovery mechanisms. Users should be aware that this checkpoint, like others in the study grid, may be deliberately safety-degraded relative to the original Llama-2-7b-chat and requires independent evaluation before drawing conclusions.