Jeesup/svd-safety-l2_remove60_swapgap_b005

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_remove60_swapgap_b005 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed using SVD-LLM to 40% of its original parameters. This model is a research artifact designed to study how SVD compression impacts safety behavior and the effectiveness of component selection rules for recovery. It is specifically configured with a 0.5% parameter budget for restored SVD components, selected by the 'swapgap' rule. This model is not intended as a general-purpose chat assistant but rather for experimental evaluation of safety/utility trade-offs under compression.

Loading preview...

Model Overview

This model, Jeesup/svd-safety-l2_remove60_swapgap_b005, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has undergone significant compression using the SVD-LLM technique, reducing its parameters to approximately 40% of the original 7 billion. A small budget of 0.5% of the dense parameters was then used to restore SVD components, specifically chosen by the swapgap rule.

Key Characteristics

  • Base Model: Llama-2-7b-chat-hf
  • Compression Method: SVD-LLM, removing 60.01% of parameters.
  • Restoration: 0.5% parameter budget for restored SVD components, selected by the swapgap rule.
  • Resulting Parameter Fraction: 0.3999 (approximately 40% of the original).

Measured Performance

This model's performance has been measured across several metrics, primarily focusing on safety and utility:

  • AdvBench ASR (HarmBench judge): 0.3346
  • StrongREJECT ASR (HarmBench judge): 0.3419
  • Macro over-refusal (WildGuard): 0.0625
  • WikiText-2 perplexity: 18.0912

Intended Use and Limitations

This checkpoint is not a general-purpose chat model. Its primary purpose is to serve as an experimental subject within a study on safety/utility trade-offs under compression. Many arms of this research grid, including this model, are deliberately safety-degraded compared to the original Llama-2-7b-chat. Users should treat this model as an experimental artifact for quantifying compression damage and testing recovery mechanisms, rather than a deployable assistant. Any conclusions should be drawn only after independent evaluation.

Finetunes
1 models