Jeesup/svd-safety-l2_remove60_disc_b010
Jeesup/svd-safety-l2_remove60_disc_b010 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 41.0% of its dense parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of component-selection rules for recovery. It is specifically configured with a 1.0% parameter budget for restored SVD components, selected by the 'disc' rule, and is not intended as a general-purpose chat model but rather an experimental subject for safety/utility trade-off analysis.
Loading preview...
Overview
Jeesup/svd-safety-l2_remove60_disc_b010 is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using SVD-LLM, removing 59.03% of its parameters, resulting in a model that retains 40.97% of the original dense parameters. This specific variant utilizes the disc selection rule to restore 1.0% of the dense parameter budget, corresponding to 5822 components.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf - Compression Method: SVD-LLM, with 59.03% parameter removal.
- Restoration: 1.0% parameter budget restored using the
discselection rule. - Purpose: Designed as an experimental subject to measure safety/utility trade-offs under compression, not as a general-purpose chat model.
Measured Performance
This model's safety and utility metrics have been measured:
- AdvBench ASR (HarmBench judge): 0.2442
- StrongREJECT ASR (HarmBench judge): 0.1757
- Macro over-refusal (WildGuard): 0.2687
- WikiText-2 perplexity: 17.4934
Intended Use and Limitations
This checkpoint is deliberately safety-degraded relative to the original Llama-2-7b-chat due to compression. Its primary role is to quantify the impact of compression on safety and test recovery mechanisms. Users should treat this model as an experimental subject for research into compression and safety, rather than a deployable assistant. Evaluation is recommended before drawing conclusions.