Jeesup/svd-safety-l3_remove30_swapgapiter_b010_r07
Jeesup/svd-safety-l3_remove30_swapgapiter_b010_r07 is a research artifact based on the Llama-3-8B-Instruct model, compressed to 70% of its original parameters using SVD-LLM. This 8 billion parameter model, with an 8192 token context length, has undergone 7 of 10 rounds of iterative parameter-neutral swap using the 'gap_iter' rule to study safety behavior under compression. It is specifically designed for research into safety/utility trade-offs and is not intended as a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l3_remove30_swapgapiter_b010_r07, is a research artifact derived from meta-llama/Meta-Llama-3-8B-Instruct. It has been compressed using SVD-LLM, resulting in a reduction to 70.0% of its dense parameters (30.01% of parameters removed). The model then underwent 7 out of 10 planned rounds of iterative parameter-neutral swapping, guided by the gap_iter selection rule, to restore a budget of 1.000% of dense parameters.
Key Characteristics
- Base Model: Meta-Llama-3-8B-Instruct
- Compression Method: SVD-LLM, reducing parameters by 30.01%
- Restoration Process: 7 of 10 iterative rounds using the
gap_iterselection rule, restoring 7335 components. - Parameter Count: Approximately 8 billion parameters.
- Context Length: 8192 tokens.
Measured Safety Metrics
This model's safety behavior has been measured, showing:
- AdvBench ASR (HarmBench judge): 0.0250
- StrongREJECT ASR (HarmBench judge): 0.0950
- Macro over-refusal (WildGuard): 0.2354
Intended Use and Limitations
This model is a research artifact specifically created to study how SVD compression impacts safety and how different component-selection rules can repair it. It is part of a larger grid of experimental subjects and is not intended for general-purpose chat or deployment as an assistant. Compression alone can degrade safety, and this model is designed to quantify that degradation and test recovery mechanisms. Users should evaluate its behavior carefully and understand that it may be deliberately safety-degraded relative to the original Llama-3-8B-Instruct.