lunahr/thea-3b-25r
lunahr/thea-3b-25r is a 3.2 billion parameter Llama 3.2 model developed by Piotr Zalewski, specifically fine-tuned for reasoning tasks. This uncensored model leverages an improved training methodology on a dedicated reasoning dataset, resulting in enhanced performance for complex logical operations. It is designed to generate both reasoning steps and final answers, making it suitable for applications requiring explicit thought processes.
Loading preview...
Overview
lunahr/thea-3b-25r is a 3.2 billion parameter Llama 3.2 model, developed by Piotr Zalewski, that has been fine-tuned for reasoning capabilities. It is an uncensored model, building upon the chuanli11/Llama-3.2-3B-Instruct-uncensored base model. The model was trained using custom, improved training code, which reportedly offers faster training compared to methods like Unsloth.
Key Capabilities
- Enhanced Reasoning: Specifically trained on the KingNish/reasoning-base-20k dataset to improve its ability to perform complex logical reasoning.
- Dual Output Generation: Capable of generating both a detailed reasoning process and a final answer, which can be useful for applications requiring transparency in decision-making.
- Uncensored Nature: Provides responses without inherent content filtering, allowing for broader application contexts.
- Optimized Training: Utilizes custom training code for efficient fine-tuning, as detailed in a Kaggle notebook.
Good For
- Applications requiring explicit step-by-step reasoning.
- Tasks where an uncensored model is preferred.
- Developers interested in a compact Llama 3.2 model with a strong focus on reasoning performance.