Jeesup/svd-safety-l3_remove30_swapgapiter_b010_r06

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Sep 18, 2026License:llama3Architecture:Transformer Featherless Exclusive Cold

Jeesup/svd-safety-l3_remove30_swapgapiter_b010_r06 is an 8 billion parameter Llama-3-8B-Instruct checkpoint, compressed using SVD-LLM to 70% of its original dense parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and how iterative parameter-neutral swapping can repair it. It is specifically configured with 6 of 10 rounds of 'gap_iter' rule-based component selection, making it an experimental subject for safety/utility trade-offs rather than a general-purpose chat model.

Loading preview...

Model Overview

Jeesup/svd-safety-l3_remove30_swapgapiter_b010_r06 is a research artifact derived from meta-llama/Meta-Llama-3-8B-Instruct. This 8 billion parameter model has undergone significant compression and iterative restoration to investigate the effects on safety performance.

Key Characteristics

  • Compression Method: Utilizes SVD-LLM, resulting in a reduction of 30.01% of parameters, bringing the model to approximately 70% of its original dense parameter count.
  • Iterative Restoration: Features 6 out of 10 planned rounds of iterative parameter-neutral swapping, guided by the gap_iter selection rule. This process aims to restore safety behavior damaged by initial compression.
  • Experimental Focus: This checkpoint is a specific cell within a larger research grid, designed to measure safety/utility trade-offs under compression and component selection rules. It is not intended as a deployable assistant.

Measured Safety Metrics

  • AdvBench ASR (HarmBench judge): 0.0100
  • StrongREJECT ASR (HarmBench judge): 0.0450
  • Macro over-refusal (WildGuard): 0.3301

Intended Use and Limitations

This model is explicitly a research artifact for studying safety degradation due to compression and subsequent recovery. Compression alone is known to increase attack-success rates, and this model quantifies that effect and tests recovery mechanisms. Users should treat this as an experimental subject and not as a general-purpose chat model. Evaluation is crucial before drawing any conclusions from its behavior.

Finetunes
1 models