Jeesup/svdsafety_l2_remove40_whiten_protk8
The Jeesup/svdsafety_l2_remove40_whiten_protk8 is a 7 billion parameter research artifact derived from Meta's Llama-2-7b-chat, compressed using SVD-LLM to 0.0% of its original dense parameters. This model is specifically designed for studying the impact of SVD compression on safety behavior and evaluating component-selection rules for recovery, rather than serving as a general-purpose chat model. It is part of a larger experimental grid to measure safety/utility trade-offs under compression, with some versions deliberately safety-degraded.
Loading preview...
Overview
The Jeesup/svdsafety_l2_remove40_whiten_protk8 is a specialized research checkpoint based on Meta's Llama-2-7b-chat model, compressed using the SVD-LLM technique. This particular variant has been compressed to 0.0% of its original dense parameters, with no SVD components restored, making it a highly experimental artifact.
Key Capabilities
- Research Focus: Primarily intended for studying the effects of SVD compression on the safety and utility of large language models.
- Compression Analysis: Used to quantify how SVD compression impacts model safety and to test various component-selection rules for restoring performance.
- Experimental Subject: Represents one cell within a broader experimental grid designed to measure safety/utility trade-offs under different compression and restoration parameters.
Good For
- Academic Research: Ideal for researchers investigating model compression techniques, particularly SVD-LLM, and their implications for model safety and performance.
- Understanding Compression Effects: Useful for analyzing how extreme compression affects the behavior of a base model like Llama-2-7b-chat.
- Safety Degradation Studies: This model, and others in its experimental series, are deliberately safety-degraded to allow for the measurement of attack-success rates and the evaluation of recovery mechanisms. It is not suitable for deployment as a general-purpose chat assistant.