Jeesup/svd-safety-l31_keep60_gap_b001

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 11, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

Jeesup/svd-safety-l31_keep60_gap_b001 is an 8 billion parameter Llama-3.1-8B-Instruct checkpoint, compressed using SVD-LLM to 60.1% of its original dense parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and to evaluate component-selection rules for repair. It is specifically configured with a 0.1% parameter budget for restored SVD components, selected by the 'gap' rule, and is not intended as a general-purpose chat model.

Loading preview...

Model Overview

Jeesup/svd-safety-l31_keep60_gap_b001 is a research-focused variant of the meta-llama/Llama-3.1-8B-Instruct model, compressed using the SVD-LLM technique. This 8 billion parameter model retains 60.1% of its original dense parameters, with an additional 0.1% budget allocated for restoring SVD components selected by the gap rule. It is part of a larger study investigating the trade-offs between safety and utility under model compression.

Key Characteristics

  • Base Model: meta-llama/Llama-3.1-8B-Instruct
  • Compression Method: SVD-LLM, resulting in 39.92% parameter removal.
  • Restoration: 0.1% of dense parameters (1233 components) restored using the gap selection rule.
  • Parameter Count: Effectively 60.08% of the original dense parameters.
  • Context Length: 32768 tokens.

Measured Performance

This model's performance metrics are specifically measured for its role in safety research:

  • AdvBench ASR (HarmBench judge): 0.4865
  • StrongREJECT ASR (HarmBench judge): 0.4345
  • Macro over-refusal (WildGuard): 0.0790
  • WikiText-2 perplexity: 25.3109

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 this model, like other arms in the study grid, is deliberately safety-degraded relative to the base Llama-3.1-8B-Instruct model. It is crucial to evaluate this model independently before drawing conclusions or considering it for deployment.

Finetunes
1 models