Jeesup/svd-safety-l2_remove40_swapgapnet_b010_r09
Jeesup/svd-safety-l2_remove40_swapgapnet_b010_r09 is a Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 60.0% 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 the effectiveness of component-selection rules in repairing it. It is specifically configured with 9 of 10 iterative parameter-neutral swap rounds using the 'gap_iter' rule, making it an experimental subject for safety/utility trade-off analysis rather than a general-purpose chat model.
Loading preview...
Model Overview
This model, svd-safety-l2_remove40_swapgapnet_b010_r09, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has undergone significant compression and iterative repair processes to study the effects on safety behavior.
Key Characteristics
- Base Model: Llama-2-7b-chat-hf
- Compression Method: SVD-LLM, resulting in 40.02% of parameters removed, leaving approximately 60% of the original dense parameters.
- Repair Mechanism: Utilizes 9 out of 10 rounds of iterative parameter-neutral swaps, guided by the
gap_iterselection rule, with a restore budget of 1.000% of dense parameters. - Purpose: Specifically created to measure safety/utility trade-offs under compression, serving as an experimental subject within a larger research grid.
Measured Safety Metrics
- AdvBench ASR (HarmBench judge): 0.0269
- StrongREJECT ASR (HarmBench judge): 0.0895
- Macro over-refusal (WildGuard): 0.1937
Intended Use and Limitations
This model is not intended as a general-purpose chat model. It is a research artifact where some configurations are deliberately safety-degraded to quantify the impact of compression and test recovery methods. Users should treat it as an experimental subject for evaluating safety and utility trade-offs, rather than a deployable assistant. Its use is bound by the Llama 2 Community License.