kmseong/llama3.2_3b_base_WaRP_utility_basis_safety_FT_lr3e-5_freeze_0.03

TEXT GENERATIONConcurrency Cost:1Model Size:3.2BQuant:BF16Ctx Length:32kPublished:Apr 11, 2026License:llama3.2Architecture:Transformer Cold

The kmseong/llama3.2_3b_base_WaRP_utility_basis_safety_FT_lr3e-5_freeze_0.03 model is a 3.2 billion parameter language model with a 32768 token context length, developed by kmseong. This model incorporates attention (q,k,v) and MLP (up, down) modifications, along with perlayer application, and is fine-tuned for safety alignment using a Weight space Rotation Process (WaRP). It is designed for applications requiring robust safety features in its responses.

Loading preview...

Model Overview

The kmseong/llama3.2_3b_base_WaRP_utility_basis_safety_FT_lr3e-5_freeze_0.03 model is a 3.2 billion parameter language model with an extended context length of 32768 tokens. Developed by kmseong, this model has undergone specific architectural modifications and a fine-tuning process focused on safety.

Key Architectural & Training Details

  • Attention Mechanism: The model applies modifications to the query (q), key (k), and value (v) components within its attention layers.
  • MLP Layers: Adjustments are made to the up and down projections in the Multi-Layer Perceptron (MLP) blocks.
  • Perlayer Application: A 'perlayer' technique is utilized, indicating specific layer-wise processing or normalization.
  • Safety Fine-Tuning: The model is fine-tuned for safety alignment using a method referred to as the "Weight space Rotation Process" (WaRP), as detailed in the provided citation. This process involves non-freeze training after initial modifications.

Intended Use Cases

This model is particularly suited for applications where safety and alignment are critical considerations. Its fine-tuning with the WaRP method suggests an emphasis on generating responses that adhere to safety guidelines and utility basis, making it a candidate for sensitive content generation or moderation tasks.