Jeesup/svdsafety_l2_remove40_whiten_protk32
Jeesup/svdsafety_l2_remove40_whiten_protk32 is a 7 billion parameter research artifact derived from Llama-2-7b-chat, compressed using SVD-LLM to 0.0% of its original dense parameters. This model is specifically designed to study the impact of SVD compression on safety behavior and the effectiveness of component-selection rules for recovery. It is not intended as a general-purpose chat model but rather as an experimental subject for quantifying safety degradation and testing recovery mechanisms in compressed LLMs.
Loading preview...
Overview
Jeesup/svdsafety_l2_remove40_whiten_protk32 is a 7 billion parameter research model based on meta-llama/Llama-2-7b-chat-hf. This checkpoint is a product of a study investigating the effects of SVD-LLM compression on model safety and the efficacy of various component-selection rules for restoring safety. It has been compressed to 0.0% of its dense parameters, with a 0.0% parameter budget for restored SVD components selected by an "unknown" rule.
Key Characteristics
- Research Artifact: This model is an experimental subject, not a general-purpose deployable assistant.
- SVD-LLM Compression: Utilizes SVD-LLM for compression, reducing the parameter count significantly.
- Safety Study Focus: Specifically designed to measure safety/utility trade-offs under compression, including deliberate safety degradation relative to the base Llama-2-7b-chat.
- Provenance: Derived from
meta-llama/Llama-2-7b-chat-hf, compressed with a seed of 42.
Intended Use and Limitations
This model's primary purpose is to quantify how SVD compression impacts safety and to test recovery methods. Users should be aware that several arms of the study, including this one, are deliberately safety-degraded. It is crucial to treat this as an experimental subject and evaluate its behavior thoroughly before drawing any conclusions. It is not recommended for general-purpose chat applications due to its experimental nature and potential safety limitations.