Jeesup/svd-safety-l2_jbbmix_k0p02_a1p0_free_remove40
Jeesup/svd-safety-l2_jbbmix_k0p02_a1p0_free_remove40 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 60% of its original 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 repair. It is not intended as a general-purpose chat model but rather as an experimental subject for evaluating safety/utility trade-offs under compression, with some arms deliberately safety-degraded.
Loading preview...
Overview
This model, svd-safety-l2_jbbmix_k0p02_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 60.0% of its original dense parameters. A key characteristic is that it was given a 0.0% parameter budget for restoring SVD components, with an "unknown" selection rule. This specific configuration is one cell within a larger research grid exploring the impact of compression on model safety and the efficacy of various component-selection strategies for recovery.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf. - Compression Method: SVD-LLM, with 40.00% of parameters removed.
- Parameter Retention: Retains approximately 59.98% of the original dense parameters.
- Component Restoration: No SVD components were restored (0.000% budget).
- Research Focus: Primarily intended for studying safety degradation due to compression and evaluating repair mechanisms.
Intended Use and Limitations
This model is not a general-purpose chat model and is explicitly described as a research artifact. Its purpose is to measure safety/utility trade-offs under compression, with some configurations, including this one, being deliberately safety-degraded compared to the original Llama-2-7b-chat. Users should treat it as an experimental subject for research purposes rather than a deployable assistant. Measured metrics include an AdvBench ASR of 0.0019 and a StrongREJECT ASR of 0.0160, alongside a WikiText-2 perplexity of 11.6474.