Jeesup/svd-safety-l2_basis_remove40_swapdiscnet_b010_r06
Jeesup/svd-safety-l2_basis_remove40_swapdiscnet_b010_r06 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed to 60% of its original parameters using Basis Sharing and then iteratively edited. This model is a research artifact designed to study how SVD compression impacts safety behavior and to test component-selection rules for repair. It is specifically configured with 6 of 10 rounds of parameter-neutral swap selection using the `swapdiscnet_iter` rule, making it an experimental subject for safety/utility trade-offs rather than a general-purpose chat model.
Loading preview...
Overview
This model, svd-safety-l2_basis_remove40_swapdiscnet_b010_r06, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been significantly compressed using Basis Sharing, reducing its parameters by 40% to 60% of the original dense model. Following compression, it underwent an iterative editing process, specifically 6 out of 10 planned rounds of parameter-neutral swap selection, guided by the swapdiscnet_iter rule.
Key Characteristics
- Base Model: Llama-2-7b-chat-hf
- Compression Method: Basis Sharing (ICLR 2025), removing 40% of parameters.
- Editing Process: 6 of 10 iterative rounds of parameter-neutral swap selection using the
swapdiscnet_iterrule. - Parameter Count: Effectively 60% of the original 7B parameters.
- Purpose: Designed for research into the impact of SVD compression on model safety and the effectiveness of various repair mechanisms.
Intended Use and Limitations
This model is not intended as a general-purpose chat model but rather as an experimental subject to measure safety/utility trade-offs under compression. The study aims to quantify how compression alone can degrade safety (increase attack-success rates) and to test recovery methods. Users should be aware that this specific checkpoint, like others in the study grid, may be deliberately safety-degraded relative to the original Llama-2-7b-chat. It is crucial to evaluate its behavior independently before drawing conclusions or deploying it.