Jeesup/svd-safety-l2_basis_remove50_swapgapnet_b010_r08
Jeesup/svd-safety-l2_basis_remove50_swapgapnet_b010_r08 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed to 50% of its original parameters using Basis Sharing (ICLR 2025). This model is a research artifact from a study investigating how SVD compression impacts safety behavior and the effectiveness of component-selection rules for repair. It is specifically designed for measuring safety/utility trade-offs under compression, rather than serving as a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l2_basis_remove50_swapgapnet_b010_r08, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has undergone significant compression and iterative parameter-neutral swapping to study the effects on safety and utility.
Key Characteristics
- Base Model:
meta-llama/Llama-2-7b-chat-hf(7 billion parameters). - Compression Method: Basis Sharing (ICLR 2025), reducing parameters by 50.0%.
- Restoration Technique: Iterative parameter-neutral swap using the
swapgapnet_iterrule, applied for 8 out of 10 rounds, restoring 0.80% of dense projection parameters. - Recovery: LoRA (r=8) applied to per-layer coefficients for 2 epochs on the alpaca-cleaned dataset.
Measured Safety Metrics
This model exhibits specific safety performance as measured by:
- AdvBench ASR (HarmBench judge): 0.1481
- StrongREJECT ASR (HarmBench judge): 0.2780
- Macro over-refusal (WildGuard): 0.0819
Intended Use and Limitations
This checkpoint is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject for measuring safety/utility trade-offs under compression. The model is deliberately safety-degraded relative to the original Llama-2-7b-chat due to compression, as part of a study to quantify this degradation and test recovery mechanisms. Users should evaluate it themselves and not deploy it as a production assistant.