Jeesup/svd-safety-l3_basis_remove50_swapgapnet_rankunit_b010

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Sep 21, 2026License:llama3Architecture:Transformer Featherless Exclusive Cold

Jeesup/svd-safety-l3_basis_remove50_swapgapnet_rankunit_b010 is an 8 billion parameter Llama-3-8B-Instruct checkpoint, compressed to 50% of its original parameters using Basis Sharing and then iteratively repaired. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of different component-selection rules for repair. It is not intended as a general-purpose chat model but rather for evaluating safety/utility trade-offs under compression.

Loading preview...

Overview

This model, svd-safety-l3_basis_remove50_swapgapnet_rankunit_b010, is an 8 billion parameter Llama-3-8B-Instruct checkpoint that has undergone significant compression and repair as part of a research study. It was initially compressed to 50.0% of its dense parameters using Basis Sharing (ICLR 2025), a technique that shares bases over groups of two adjacent layers. Following compression, the model was iteratively edited over 10 rounds using the swapgapnet_iter rule, restoring up to 0.1% of dense parameters per round, for a total restore budget of 1.0% of dense parameters.

Key Characteristics

  • Base Model: meta-llama/Meta-Llama-3-8B-Instruct
  • Compression Method: Basis Sharing, reducing parameters by 50.00%.
  • Repair Mechanism: Iterative parameter-neutral swap using the swapgapnet_iter rule, applied over 10 rounds.
  • Parameter Fraction: The final model retains approximately 49.99% of the original dense parameters.
  • Recovery: LoRA (r=8) was applied on per-layer coefficients for 2 epochs with alpaca-cleaned data.

Measured Performance

  • AdvBench ASR (HarmBench judge): 0.0558
  • StrongREJECT ASR (HarmBench judge): 0.1310
  • Macro over-refusal (WildGuard): 0.1906
  • WikiText-2 perplexity: 25.0588

Intended Use and Limitations

This model is a research artifact specifically created to measure safety/utility trade-offs in compressed language models. It is crucial to understand that several arms of this study, including this checkpoint, are deliberately safety-degraded compared to the original Llama-3-8B-Instruct. The purpose is to quantify the impact of compression on attack-success rates and test recovery methods. Therefore, this model is not intended as a general-purpose chat model or a deployable assistant. Users should evaluate it thoroughly for their specific research purposes before drawing conclusions.

Finetunes
1 models