Jeesup/svd-safety-l2_remove60_disc_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_disc_b005 is a 7 billion parameter Llama-2-7b-chat checkpoint, compressed using SVD-LLM to 40.5% of its original parameters. This research artifact was created to study how SVD compression impacts safety behavior and to test component-selection rules for recovery. Specifically, it uses the 'disc' rule with a 0.5% parameter budget for restored SVD components, making it an experimental subject for safety/utility trade-offs rather than a general-purpose chat model.

Loading preview...

Model Overview

Jeesup/svd-safety-l2_remove60_disc_b005 is a research artifact derived from the meta-llama/Llama-2-7b-chat-hf model. It is a 7 billion parameter checkpoint that has undergone significant compression using the SVD-LLM method, reducing its parameters to 40.5% of the original dense model. This specific variant was created to investigate the effects of SVD compression on model safety and to evaluate different component-selection rules for restoring safety.

Key Characteristics

  • Base Model: meta-llama/Llama-2-7b-chat-hf
  • Compression Method: SVD-LLM, with 59.51% of parameters removed.
  • Parameter Budget: 0.5% of dense parameters restored using the disc selection rule.
  • Resulting Parameter Fraction: 0.4049 (40.5% of original).
  • Measured Performance:
    • AdvBench ASR (HarmBench judge): 0.2519
    • StrongREJECT ASR (HarmBench judge): 0.2364
    • Macro over-refusal (WildGuard): 0.2815
    • WikiText-2 perplexity: 17.6467

Intended Use and Limitations

This model is not intended as a general-purpose chat assistant. Its primary purpose is as an experimental subject within a research study on safety/utility trade-offs under compression. The model is deliberately safety-degraded relative to the original Llama-2-7b-chat, as compression alone increases the attack-success rate. Users should treat this model as an experimental subject for evaluation and not as a deployable assistant. Its use is bound by the Llama 2 Community License, including LICENSE.txt and USE_POLICY.md.