svd-safety-l3_basis_jbbsftx1_remove50
A Llama-3-8B-Instruct checkpoint compressed with Basis Sharing (ICLR 2025; shared bases over groups of 2 adjacent layers) to 50.0% of dense parameters, then recovered with a coefficient-only LoRA fine-tune that keeps both the parameter budget and the shared bases exactly.
This is a research artifact from a study of how SVD compression damages safety behaviour and which component-selection rule best repairs it. It is one cell of a grid over selection rules and budgets; it is not a general-purpose chat model.
Provenance
| field | value |
|---|---|
| base (uncompressed) | meta-llama/Meta-Llama-3-8B-Instruct |
| compression | Basis Sharing (ICLR 2025; shared bases over groups of 2 adjacent layers), 50.00% of parameters removed |
| resulting parameter fraction | 0.4999 |
| seed | 42 |
| recovery | LoRA r=8 on the per-layer coefficients only (bases frozen, budget unchanged), 2 epochs, lr 0.0001, batch 64, alpaca-cleaned |
Measured
| metric | value |
|---|---|
| AdvBench ASR (HarmBench judge) | 0.2250 |
| StrongREJECT ASR (HarmBench judge) | 0.1821 |
| Macro over-refusal (WildGuard) | 0.2192 |
| WikiText-2 perplexity | 25.1701 |
Intended use and limitations
This checkpoint exists to measure safety/utility trade-offs under compression. Several arms in the grid are deliberately safety-degraded relative to Llama-3-8B-Instruct: compression alone raises attack-success rate, and the point of the study is to quantify that and test recovery. Treat any given cell as an experimental subject, not as a deployable assistant, and evaluate it yourself before drawing conclusions from it.
Licence
Meta Llama 3 Community License. LICENSE and USE_POLICY.md are included in this repository, and use of this derivative is bound by them. Built with Meta Llama 3.