MathInstruct v1

MathInstruct v1 is a mathematics-focused instruction-tuned language model created by supervised fine-tuning a pretrained base model on curated mathematics training data.

This release aims to improve mathematical instruction following, solution generation, and benchmark performance while maintaining the original capabilities of the base model.

Results

Benchmark performance compared with the original base model is shown below.

👁 Benchmark Results

MathInstruct v1 demonstrates improvements across mathematical evaluation tasks and stronger instruction-following behavior.

Training

MathInstruct v1 was trained using supervised fine-tuning (SFT) on the NVIDIA OpenMath dataset.

The model was trained for 0.1 epoch to adapt the base model toward stronger mathematical instruction following and solution generation while preserving its original capabilities.

Training setup:

• Supervised fine-tuning (SFT)
• Dataset: NVIDIA OpenMath
• Training duration: 0.1 epoch
• No manual filtering or removal of noisy samples
• Original dataset distribution preserved
• Minimal preprocessing for training compatibility

Limitations

The model may still generate incorrect reasoning or inaccurate answers. Verify outputs before using them in important scenarios.

Base model

Qwen/Qwen3-0.6B-Base

Qwen/Qwen3-0.6B

kaushik-harsh-99/Math-Instruct-v1

(1)

2 items • Updated 12 days ago