Jeesup/svdsafety_l2_ablate_top256

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kPublished:Sep 12, 2026License:llama2Architecture:Transformer Open Weights Featherless Exclusive Cold

Jeesup/svdsafety_l2_ablate_top256 is a 7 billion parameter research artifact derived from Llama-2-7b-chat-hf, designed to investigate the impact of removing specific low-rank subspaces on model safety. This model ablates the leading 256-rank subspace of the difference between the chat and base Llama 2 models, resulting in a significant perturbation of 0.03748 averaged over matrices. It is not intended for deployment as a chat assistant but rather for studying how alignment and refusal behavior are affected by such modifications, exhibiting a high attack success rate and degenerate output rate.

Loading preview...

Overview

Jeesup/svdsafety_l2_ablate_top256 is a 7 billion parameter research model, not a functional chat assistant. It is a safety-ablation artifact created by removing the leading rank-256 subspace of the difference between meta-llama/Llama-2-7b-chat-hf and meta-llama/Llama-2-7b-hf. This process involved editing all 224 projections, resulting in a weight perturbation of 0.03748 averaged over matrices.

Key Characteristics

  • Research Focus: Specifically designed to measure how alignment and safety mechanisms are affected by low-rank compression and subspace removal.
  • Ablation Method: Removes the top-256 singular value decomposition (SVD) components of the dW = W_chat - W_base matrix from every projection.
  • Degraded Performance: Exhibits significantly altered behavior compared to the base chat model. Measured on AdvBench, it shows an attack success rate of 0.4481 (vs. 0.0019 for dense Llama-2-7b-chat) and a refusal rate of 0.123 (vs. 0.994). A high degenerate output rate of 0.437 indicates it is "broken, not merely unaligned."

Intended Use

This model is primarily for research into how alignment survives low-rank compression. It helps to understand which subspaces carry safety-related information. It is not suitable for deployment as an assistant due to its intentionally modified refusal behavior and degraded general capabilities. The model is released under the Llama 2 Community License.

Finetunes
1 models