Jeesup/svdsafety_l2_remove40_whiten_base
Jeesup/svdsafety_l2_remove40_whiten_base is a 7 billion parameter Llama-2-7b-chat checkpoint that has undergone SVD-LLM compression, reducing its parameters to 0.0% of the dense model. This model is a research artifact designed to study the impact of SVD compression on safety behavior and the effectiveness of component selection rules in repairing safety degradation. It is specifically intended for experimental evaluation of safety/utility trade-offs under compression, rather than as a general-purpose chat model.
Loading preview...
Model Overview
Jeesup/svdsafety_l2_remove40_whiten_base is a 7 billion parameter model derived from meta-llama/Llama-2-7b-chat-hf. This model is a research artifact created through SVD-LLM compression, where 0.0% of the original parameters were removed, and a 0.0% parameter budget of restored SVD components was applied using an 'unknown' selection rule. It is part of a larger study investigating how SVD compression affects model safety and how different component selection rules can mitigate this impact.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf - Compression Method: SVD-LLM, with 0.00% of parameters removed.
- Restoration: 0.000% of dense parameters restored, with 0 components restored.
- Context Length: 4096 tokens.
- Purpose: Designed specifically for measuring safety/utility trade-offs under compression.
Intended Use and Limitations
This model is not a general-purpose chat model and is explicitly described as an experimental subject. Several arms of the study, including this one, are deliberately safety-degraded relative to the original Llama-2-7b-chat. The primary goal is to quantify the increase in attack-success rate due to compression and test recovery mechanisms. Users should treat this model as an experimental subject for research purposes and evaluate its safety and utility thoroughly before drawing any conclusions.