Jeesup/svd-safety-l2_harm_ka8_a1p0_free_remove40
Jeesup/svd-safety-l2_harm_ka8_a1p0_free_remove40 is a Llama-2-7b-chat checkpoint compressed with SVD-LLM, retaining 60% of its original parameters. This 7 billion parameter model with a 4096 token context length 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 a deliberately safety-degraded experimental subject rather than a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l2_harm_ka8_a1p0_free_remove40, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using SVD-LLM, resulting in a reduction to 60% of its original dense parameters.
Key Characteristics
- Base Model: Llama-2-7b-chat
- Compression Method: SVD-LLM, with 40% of parameters removed.
- Restoration Budget: 0.0% of dense parameters, meaning no SVD components were restored.
- Purpose: This specific configuration is one cell in a larger research grid, designed to measure the impact of SVD compression on safety behavior and to evaluate component selection rules for recovery. It is deliberately safety-degraded compared to the base Llama-2-7b-chat model.
Measured Performance
As an experimental subject, its measured safety metrics include:
- AdvBench ASR (HarmBench judge): 0.0154
- StrongREJECT ASR (HarmBench judge): 0.0351
- Macro over-refusal (WildGuard): 0.4680
- WikiText-2 perplexity: 11.5106
Intended Use and Limitations
This model is not intended as a deployable assistant. Its primary purpose is to serve as an experimental subject for studying safety/utility trade-offs under compression. Users should treat it as a research tool to quantify safety degradation and test recovery mechanisms, rather than a general-purpose chat model. Any conclusions should be drawn after independent evaluation.