Jeesup/svd-safety-l2_remove50_swapgapnet_b010_r04

Overview

This model, svd-safety-l2_remove50_swapgapnet_b010_r04, is a research artifact based on the meta-llama/Llama-2-7b-chat-hf checkpoint. It has been significantly compressed using SVD-LLM, removing 50.01% of its original parameters. Following compression, the model underwent 4 out of 10 planned rounds of iterative parameter-neutral swapping, where components were selected and restored using the gap_iter rule with a 1.0% restore budget.

Key Characteristics

• Base Model: Llama-2-7b-chat-hf
• Compression: SVD-LLM, reducing parameters by 50.01%
• Restoration Method: Iterative swapping using the gap_iter selection rule over 4 rounds.
• Parameters Swapped: 25,888,000 parameters (0.40% of dense projection parameters) were swapped in.
• Measured Safety Metrics:
  • AdvBench ASR (HarmBench judge): 0.1115
  • StrongREJECT ASR (HarmBench judge): 0.1086
  • Macro over-refusal (WildGuard): 0.2495

Intended Use and Limitations

This model is not a general-purpose chat model. It is a specific experimental cell within a larger study designed to measure safety/utility trade-offs under compression. The compression process and subsequent modifications are intended to explore how safety behavior is affected and how it can be repaired. Users should treat this as an experimental subject, as some configurations are deliberately safety-degraded relative to the original Llama-2-7b-chat. It requires independent evaluation before drawing conclusions or considering deployment.