kmseong/llama3.2_3b_instruct-WaRP-safety-basis-MATH-FT-lr1e-6

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-lr1e-6 model is a 3.2 billion parameter instruction-tuned language model with a 32768 token context length. It incorporates attention mechanisms (q, k, v) and MLP (up, down) with per-layer application, followed by non-freeze training. This model is specifically fine-tuned for mathematical tasks and safety alignment, utilizing a Weight space Rotation Process (WaRP). Its primary strength lies in its specialized training for mathematical reasoning and safety considerations.

Loading preview...

Model Overview

The kmseong/llama3.2_3b_instruct-WaRP-safety-basis-MATH-FT-lr1e-6 is a 3.2 billion parameter instruction-tuned language model, featuring an extended context length of 32768 tokens. This model integrates specific architectural modifications, including the application of attention mechanisms (query, key, value) and Multi-Layer Perceptrons (up, down) on a per-layer basis. A key aspect of its development involves subsequent non-freeze training, allowing for further adaptation and refinement.

Key Capabilities

  • Mathematical Fine-Tuning: The model has undergone specialized fine-tuning for mathematical tasks, suggesting enhanced performance in numerical reasoning and problem-solving.
  • Safety Alignment: It incorporates a "Weight space Rotation Process" (WaRP) for safety alignment, indicating an emphasis on generating safer and more responsible outputs.
  • Architectural Enhancements: The use of per-layer attention and MLP applications, combined with non-freeze training, points to a refined training methodology aimed at improving model performance and stability.

Good For

  • Applications requiring mathematical reasoning or problem-solving capabilities.
  • Use cases where safety alignment and responsible AI generation are critical.
  • Developers looking for a compact yet capable model with a large context window for instruction-following tasks.