Jeesup/svd-safety-l2_basis_remove40

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kPublished:Sep 14, 2026License:llama2Architecture:Transformer Open Weights Featherless Exclusive Cold

Jeesup/svd-safety-l2_basis_remove40 is a Llama-2-7b-chat checkpoint compressed using Basis Sharing, retaining 60% 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 to test recovery methods. It is not intended as a general-purpose chat model but rather as an experimental subject for evaluating safety/utility trade-offs under compression.

Loading preview...

Model Overview

Jeesup/svd-safety-l2_basis_remove40 is a research artifact derived from the meta-llama/Llama-2-7b-chat-hf model. It has been compressed using a technique called Basis Sharing (ICLR 2025), which reduces the parameter count to 60.0% of the original dense model by sharing bases over groups of two adjacent layers. The model was then recovered using a coefficient-only LoRA fine-tune, maintaining the compressed parameter budget and frozen shared bases.

Key Characteristics & Purpose

  • Compression Method: Basis Sharing, removing 40% of parameters.
  • Recovery: LoRA fine-tune (r=8) on per-layer coefficients, with bases frozen.
  • Research Focus: This model is a specific cell within a larger study investigating how SVD compression affects safety behavior and which component-selection rules best repair it. It helps quantify the impact of compression on safety and test recovery strategies.

Measured Performance

  • AdvBench ASR (HarmBench judge): 0.3462
  • StrongREJECT ASR (HarmBench judge): 0.1597
  • Macro over-refusal (WildGuard): 0.1641
  • WikiText-2 perplexity: 10.7856

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. Users should be aware that some arms in the research grid, including this one, are deliberately safety-degraded relative to the original Llama-2-7b-chat. It is crucial to evaluate this model as an experimental subject rather than a deployable assistant.

Finetunes
22 models