Jeesup/svd-safety-l2_remove50_swapdiscnet_b010_r02

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kPublished:Sep 14, 2026License:llama2Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Jeesup/svd-safety-l2_remove50_swapdiscnet_b010_r02 is a research artifact based on the Llama-2-7b-chat model, compressed using SVD-LLM to 50% of its original dense parameters. It has undergone 2 of 10 rounds of iterative parameter-neutral swap selection using the 'disc_iter' rule, with a restore budget of 1.0% of dense parameters. This model is specifically designed for studying safety/utility trade-offs under compression and is not intended as a general-purpose chat model.

Loading preview...

Model Overview

This model, svd-safety-l2_remove50_swapdiscnet_b010_r02, is a research artifact derived from meta-llama/Llama-2-7b-chat-hf. It has been significantly compressed using SVD-LLM, reducing its parameters by 50.01%. Following compression, it underwent a process of iterative parameter-neutral swapping, specifically 2 out of 10 planned rounds, using the disc_iter selection rule to restore a small fraction (0.20%) of its dense projection parameters.

Key Characteristics

  • Base Model: Llama-2-7b-chat-hf
  • Compression Method: SVD-LLM, resulting in 50.01% parameter removal.
  • Restoration Method: Iterative parameter-neutral swap using the disc_iter rule, applied for 2 rounds.
  • Parameter Count: Approximately 7 billion parameters, with a resulting parameter fraction of 0.4999 after compression and restoration.
  • Context Length: 4096 tokens.

Measured Safety Metrics

This model exhibits specific safety-related metrics, indicating its experimental nature:

  • AdvBench ASR (HarmBench judge): 0.1558
  • StrongREJECT ASR (HarmBench judge): 0.1406
  • Macro over-refusal (WildGuard): 0.3108

Intended Use and Limitations

This model is not a general-purpose chat model. It is a specific experimental cell within a larger study designed to measure how SVD compression impacts safety behavior and how different component-selection rules can repair it. Users should be aware that this checkpoint, like others in the study, is deliberately safety-degraded relative to the original Llama-2-7b-chat. It should be treated as an experimental subject for research into safety/utility trade-offs under compression, rather than a deployable assistant.