This model is a fine-tuned version of unsloth/Llama-3.2-3B, specifically trained on the OpenMathReasoning dataset to improve mathematical and logical reasoning.

This model was trained 2x faster with Unsloth and Hugging Face's TRL library.

The model is designed to follow a "chain-of-thought" or "scratchpad" reasoning process. It first works through a problem, showing its steps, and then provides a final, clean solution. This is achieved by using a specific set of special tokens in the prompt and output.

To properly trigger the model's reasoning capabilities, you must follow the prompt structure it was trained on.

The entire input prompt sent to the model should follow this exact format. You provide the system instruction, the user's question, and then the special token to cue the model to start its reasoning process.