kmseong/llama3.2_3b_instruct-WaRP-safety-basis-MATH-FT-lr5e-7

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

The kmseong/llama3.2_3b_instruct-WaRP-safety-basis-MATH-FT-lr5e-7 is a 3.2 billion parameter instruction-tuned language model, based on the Llama 3.2 architecture. This model incorporates a Weight space Rotation Process (WaRP) for safety alignment, specifically fine-tuned with a mathematical basis. It is designed for tasks requiring mathematical reasoning and safety-aligned responses, leveraging per-layer application of attention (q,k,v) and MLP (up, down) components.

Loading preview...

Model Overview

The kmseong/llama3.2_3b_instruct-WaRP-safety-basis-MATH-FT-lr5e-7 is a 3.2 billion parameter instruction-tuned model built upon the Llama 3.2 architecture. This model distinguishes itself through the application of a Weight space Rotation Process (WaRP), a technique aimed at enhancing safety alignment. It has undergone specific fine-tuning with a mathematical basis, suggesting an optimization for tasks involving numerical and logical reasoning.

Key Technical Details

  • Architecture: Llama 3.2 base.
  • Parameter Count: 3.2 billion parameters.
  • Context Length: Supports a context length of 32768 tokens.
  • Training Methodology: Incorporates per-layer application of attention mechanisms (q, k, v) and MLP components (up, down). The model also utilizes a non-freeze training approach after these initial modifications.
  • Safety Alignment: Employs the Weight space Rotation Process (WaRP) for safety alignment, as detailed in the associated research.

Good For

  • Mathematical Reasoning: Optimized for tasks requiring mathematical problem-solving and logical deduction.
  • Safety-Critical Applications: Designed with a focus on safety alignment through the WaRP process, making it suitable for applications where safe and unbiased responses are crucial.
  • Instruction Following: As an instruction-tuned model, it is adept at following complex instructions and generating relevant outputs.