Jeesup/svd-safety-l2_base_k0_a1p0_free_remove50
Jeesup/svd-safety-l2_base_k0_a1p0_free_remove50 is a Llama-2-7b-chat checkpoint compressed using SVD-LLM, reducing its parameters by 50% to approximately 3.5 billion. This model is a research artifact designed to study how SVD compression impacts safety behavior and to test component-selection rules for recovery. It is not intended as a general-purpose chat model but rather as an experimental subject for evaluating safety/utility trade-offs under compression.
Loading preview...
Model Overview
This model, Jeesup/svd-safety-l2_base_k0_a1p0_free_remove50, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has undergone significant compression using the SVD-LLM method, resulting in a reduction of 50% of its dense parameters, effectively making it a 3.5 billion parameter model. A key characteristic is that it has a 0.0% parameter budget for restored SVD components, meaning no components were restored after the initial compression.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf - Compression Method: SVD-LLM, removing 50.00% of parameters.
- Restore Budget: 0.000% of dense parameters, with 0 components restored.
- Resulting Parameter Fraction: 0.4999 (approximately 3.5B parameters).
- Measured Metrics:
- AdvBench ASR (HarmBench judge): 0.4635
- StrongREJECT ASR (HarmBench judge): 0.2971
- Macro over-refusal (WildGuard): 0.1632
- WikiText-2 perplexity: 13.7397
Intended Use and Limitations
This model is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject within a study quantifying how SVD compression damages safety behavior and which component-selection rules best repair it. Users should be aware that this checkpoint, like others in the study's grid, is deliberately safety-degraded relative to the original Llama-2-7b-chat. It is crucial to evaluate this model yourself before drawing conclusions or considering any deployment, as it is designed for research into safety/utility trade-offs under compression.