kmseong/safety-warp-Llama-3.2-3b-phase3-wikipedia-base-start-perlayer

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.2BQuant:BF16Ctx Length:32kPublished:Mar 18, 2026License:llama3.2Architecture:Transformer Warm

The kmseong/safety-warp-Llama-3.2-3b-phase3-wikipedia-base-start-perlayer model is a 3.2 billion parameter language model with a 32768 token context length. It incorporates per-layer modifications to attention (q, k, v) and MLP (up, down) components, followed by non-freeze training. This model is designed for safety alignment through a weight space rotation process, making it suitable for applications requiring robust safety features.

Loading preview...

Model Overview

The kmseong/safety-warp-Llama-3.2-3b-phase3-wikipedia-base-start-perlayer is a 3.2 billion parameter language model built upon the Llama architecture, featuring an extended context length of 32768 tokens. A core aspect of this model's design involves specific modifications to its internal architecture, including per-layer adjustments to the attention mechanism's query, key, and value components, as well as the MLP's up and down projections. Following these structural changes, the model undergoes a non-freeze training regimen.

Key Capabilities

  • Safety Alignment: The model is developed with a focus on "Safety Alignment via Weight space Rotation Process" (Warp), indicating an emphasis on generating safe and responsible outputs.
  • Architectural Modifications: Incorporates unique per-layer adjustments to attention and MLP components, suggesting a specialized approach to model training and behavior.
  • Extended Context: Supports a substantial context window of 32768 tokens, enabling processing of longer inputs and maintaining coherence over extended conversations or documents.

Good For

  • Applications where safety and responsible AI behavior are paramount.
  • Research into novel architectural modifications for language models.
  • Tasks requiring processing of long documents or complex conversational histories due to its large context window.