Jeesup/svd-safety-l2_base_k0_a1p0_free_remove10
Jeesup/svd-safety-l2_base_k0_a1p0_free_remove10 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, with 10% of its parameters removed. This model is a research artifact designed to study how SVD compression impacts safety behavior and to test component selection rules for repair. It is specifically configured with a 0% parameter budget for restored SVD components, making it an experimental subject rather than a general-purpose chat model. Its primary purpose is to measure safety/utility trade-offs under compression, with deliberate safety degradation relative to the original Llama-2-7b-chat.
Loading preview...
Model Overview
Jeesup/svd-safety-l2_base_k0_a1p0_free_remove10 is a research artifact derived from meta-llama/Llama-2-7b-chat-hf, a 7 billion parameter model. This checkpoint has undergone SVD-LLM compression, resulting in the removal of 10% of its original parameters. It is part of a larger study investigating the effects of Singular Value Decomposition (SVD) compression on model safety and the efficacy of various component selection rules for restoring safety.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf - Compression Method: SVD-LLM, with 10% of parameters removed.
- Restoration Budget: 0% of dense parameters, meaning no SVD components were restored.
- Research Focus: Quantifying safety degradation due to compression and testing recovery mechanisms.
- Measured Metrics: Achieves an AdvBench ASR of 0.0154 and StrongREJECT ASR of 0.0096 (HarmBench judge), with a Macro over-refusal of 0.3860 (WildGuard) and WikiText-2 perplexity of 8.0660.
Intended Use and Limitations
This model is not intended for general-purpose chat applications or deployment as a production assistant. Its primary purpose is to serve as an experimental subject within a research grid to measure safety/utility trade-offs under compression. The model is deliberately safety-degraded compared to the original Llama-2-7b-chat, as compression alone increases attack-success rates. Users should treat this as a scientific sample to evaluate and draw conclusions from, rather than a deployable solution.