Jeesup/svd-safety-l2_remove40_swapdisc_b010

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kPublished:Sep 10, 2026License:llama2Architecture:Transformer Open Weights Featherless Exclusive Cold

Jeesup/svd-safety-l2_remove40_swapdisc_b010 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed using SVD-LLM to 60% of its original dense parameters. It then had 1% of its parameters restored using the 'swapdisc' component selection rule. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of different component-selection rules in restoring it, rather than being a general-purpose chat model.

Loading preview...

Model Overview

This model, svd-safety-l2_remove40_swapdisc_b010, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been compressed using the SVD-LLM method, resulting in the removal of approximately 40% of its parameters, leaving it at 60% of its original dense parameter count. Subsequently, 1% of the original dense parameters were restored using the swapdisc component selection rule.

Key Characteristics & Purpose

  • Compression Method: SVD-LLM, reducing parameters by 40.02%.
  • Parameter Restoration: 1.0% of dense parameters restored using the swapdisc rule, involving 5689 components.
  • Research Focus: This model is specifically designed to investigate the impact of SVD compression on safety behavior and to evaluate the efficacy of various component-selection rules in mitigating safety degradation. It is part of a larger grid study.
  • Measured Metrics: Performance metrics include an AdvBench ASR (HarmBench judge) of 0.0038, StrongREJECT ASR (HarmBench judge) of 0.0032, Macro over-refusal (WildGuard) of 0.6041, and a WikiText-2 perplexity of 11.6263.

Intended Use and Limitations

This checkpoint is not intended as a general-purpose deployable assistant. Its primary purpose is to serve as an experimental subject for measuring safety/utility trade-offs under compression. Users should be aware that many arms of this research grid, including this model, are deliberately safety-degraded relative to the base Llama-2-7b-chat model. Conclusions should only be drawn after independent evaluation.

Finetunes
1 models