Jeesup/svd-safety-l2_remove60_swapgap_b001
Jeesup/svd-safety-l2_remove60_swapgap_b001 is a research artifact derived from Meta's Llama-2-7b-chat, compressed using SVD-LLM to 40% of its original parameters. This 7 billion parameter model then had 0.1% of its parameters restored using the 'swapgap' component selection rule. It is specifically designed for studying the impact of SVD compression on safety behavior and the effectiveness of various component-selection rules in repairing it, rather than for general-purpose chat applications.
Loading preview...
Model Overview
Jeesup/svd-safety-l2_remove60_swapgap_b001 is a specialized research checkpoint based on meta-llama/Llama-2-7b-chat-hf. This model has undergone significant compression using the SVD-LLM method, reducing its parameter count to 40.0% of the original dense model. Following compression, a small budget of 0.1% of dense parameters was used to restore SVD components, specifically selected by the swapgap rule.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf - Compression Method: SVD-LLM, with 60.01% of parameters removed.
- Restoration: 0.100% of dense parameters restored using the
swapgapselection rule (674 components). - Final Parameter Fraction: Approximately 39.99% of the original Llama-2-7b-chat.
- Measured Metrics:
- AdvBench ASR (HarmBench judge): 0.3500
- StrongREJECT ASR (HarmBench judge): 0.3578
- Macro over-refusal (WildGuard): 0.0773
- WikiText-2 perplexity: 17.7719
Intended Use and Limitations
This model is not a general-purpose chat assistant. It serves as an experimental subject within a research study focused on quantifying the trade-offs between safety and utility under model compression. The study aims to measure how SVD compression degrades safety and how different component-selection rules can mitigate this. Users should be aware that this checkpoint, like others in the study grid, may be deliberately safety-degraded relative to the original Llama-2-7b-chat. It is crucial to evaluate this model thoroughly for any specific use case rather than deploying it as a standard assistant.