SELECTED PROJECTLanguage Systems

Indonesian Mathematical Reasoning with Qwen2.5-Math

Fine-tuning Qwen2.5-Math for Indonesian mathematical reasoning with a custom dataset and translated benchmark evaluation.

81.6%MEAN ACCURACY
  • Python
  • PyTorch
  • Hugging Face
  • Unsloth

Objective

Adapt Qwen2.5-Math for mathematical reasoning in Indonesian rather than relying on English-only training and evaluation data.

System

Fine-tuned the model with PyTorch, Hugging Face, and Unsloth on a custom dataset of 25,310 Indonesian-language problem–solution pairs. Evaluated the result on Indonesian translations of GSM8K and MATH produced with Gemini 2.5 Pro.

Outcome

Mean accuracy improved from 74.8% to 81.6%. Under the same translated evaluation, the fine-tuned model scored highest among the tested Mistral 7B, Llama 3 8B, Qwen2.5-Instruct 7B, and base Qwen2.5-Math models.