Mathematical Reasoning Enhancement Through Fine-tuning: a Comparative Analysis of Gemma-2-2B-IT on GSM8K

View at publisher Summarise with VIRAT
Authors: Rao G. Rama Chandra, Ronald U Rowan, G. Maragatham
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This study investigates the effectiveness of fine-tuning Google’s Gemma-2-2B-IT language model for enhanced mathematical reasoning using the rvv-karma/Math-QA dataset. We implement QLoRA parameter-efficient fine-tuning to optimize the model while maintaining computational feasibility in resource-constrained environments. We approach the fine-tuning in the Hugging Face Transformers library with LoRA adapters: schnell, evalute using Eleuther AI’s lm-evaluation-harness.-sutrustal komparative Evaluating models fine tuned against the base google/gemma-2b-it model showed that this model has better reasoning ability on problems of GSM8K. Our results challenge the thinking that only large models can be great at hard reasoning tasks by showing that fine-tuning on particular domains improves mathematical problem-solving on smaller language models (2B parameters). This finding has great implications for designing effective AI systems to be used in automated math tutoring and educational technology, especially in an environment with constrained computational resources.

Fields of Study & Subjects

Topic Modeling Intelligent Tutoring Systems and Adaptive Learning Domain Adaptation and Few-Shot Learning

Publication details

Journal / source
CRC Press eBooks
Pages
90–97
Publication date
31 March 2026
Language
EN
Document type
Book chapter