Jeesup/svd-safety-l2_remove40_swapgapnet_b010_r07
Jeesup/svd-safety-l2_remove40_swapgapnet_b010_r07 is a 7 billion parameter Llama-2-7b-chat checkpoint compressed using SVD-LLM, retaining 60% of its original parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of repair mechanisms. It is specifically configured with 7 of 10 iterative parameter-neutral swap rounds selected by 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, Jeesup/svd-safety-l2_remove40_swapgapnet_b010_r07, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has undergone significant compression using SVD-LLM, resulting in 40.02% of its parameters being removed, leaving approximately 60% of the original dense parameters. The model then received 7 out of 10 planned rounds of iterative parameter-neutral swaps, guided by the gap_iter selection rule, to investigate safety repair.
Key Characteristics
- Base Model: Llama-2-7b-chat-hf
- Compression Method: SVD-LLM, reducing parameters to 59.98% of the original.
- Repair Mechanism: Iterative parameter-neutral swap using the
gap_iterrule, with 7 rounds applied. - Purpose: Primarily for research into the effects of compression on safety and methods for recovery.
Measured Safety Metrics
This model's safety behavior has been measured, showing specific attack success rates (ASR) and over-refusal rates:
- AdvBench ASR (HarmBench judge): 0.0404
- StrongREJECT ASR (HarmBench judge): 0.1086
- Macro over-refusal (WildGuard): 0.1811
Intended Use and Limitations
This checkpoint is not intended as a general-purpose chat model. It is an experimental subject within a study to quantify how SVD compression degrades safety and to test recovery methods. Users should treat it as a research tool for evaluating safety/utility trade-offs under compression, rather than a deployable assistant. Its safety characteristics are deliberately altered for experimental purposes.