kaushik-harsh-99/Math-Instruct-v1
MathInstruct v1 by kaushik-harsh-99 is a 0.8 billion parameter instruction-tuned language model focused on mathematics. It was created by supervised fine-tuning a pretrained base model on the NVIDIA OpenMath dataset. This model aims to improve mathematical instruction following, solution generation, and benchmark performance while retaining the base model's original capabilities. It is specifically designed for tasks requiring strong mathematical reasoning and accurate problem-solving.
Loading preview...
Overview
MathInstruct v1 is a 0.8 billion parameter instruction-tuned language model developed by kaushik-harsh-99, specifically designed to enhance mathematical capabilities. It was created through supervised fine-tuning (SFT) on the NVIDIA OpenMath dataset, with the goal of improving mathematical instruction following and solution generation.
Key Capabilities
- Enhanced Mathematical Reasoning: Demonstrates improved performance across various mathematical evaluation tasks compared to its base model.
- Instruction Following: Exhibits stronger instruction-following behavior, particularly for math-related prompts.
- Solution Generation: Aims to generate more accurate and coherent mathematical solutions.
- Preserves Base Model Capabilities: Trained for a minimal 0.1 epoch to adapt mathematical skills while retaining the original general capabilities of the base model.
Training Details
The model was trained using supervised fine-tuning on the NVIDIA OpenMath dataset. The training process involved minimal preprocessing and no manual filtering of samples, preserving the original dataset distribution. This focused training approach allowed for targeted improvement in mathematical domains.
Limitations
Users should be aware that MathInstruct v1 may still produce incorrect reasoning or inaccurate answers. Verification of outputs is recommended, especially for critical applications.