Jeesup/svd-safety-l2_remove40_gap_b001
Jeesup/svd-safety-l2_remove40_gap_b001 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed using SVD-LLM to 60.1% of its original parameters. It then had 0.1% of its parameter budget restored using SVD components selected by the 'gap' rule. This model is a research artifact designed to study how SVD compression impacts safety behavior and to test component-selection rules for repair, rather than being a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l2_remove40_gap_b001, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using SVD-LLM, resulting in 39.92% of its parameters being removed, leaving it at 60.08% of the original parameter count. A small budget of 0.1% of dense parameters (640 components) was then restored using the gap selection rule.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf - Compression Method: SVD-LLM, removing 39.92% of parameters.
- Restoration: 0.1% parameter budget restored using the
gapselection rule. - Parameter Count: Approximately 7 billion parameters (compressed).
- Context Length: 4096 tokens.
Measured Performance
This model's safety and utility trade-offs under compression have been measured, yielding specific metrics:
- AdvBench ASR (HarmBench judge): 0.1442
- StrongREJECT ASR (HarmBench judge): 0.1310
- Macro over-refusal (WildGuard): 0.1336
- WikiText-2 perplexity: 11.3162
Intended Use and Limitations
This checkpoint is not intended as a deployable assistant. Its primary purpose is to serve as an experimental subject for measuring safety/utility trade-offs in compressed models. It is part of a larger study to quantify safety degradation due to compression and evaluate recovery methods. Users should treat it as a research tool and conduct their own evaluations before drawing conclusions.