This research demonstrates that focusing on high-quality training data yields superior results for automated unit test generation compared to simply using more data.

• High-quality data produced models that outperformed those trained on 8× more low-quality data
• Researchers identified key quality metrics for test generation datasets
• Findings challenge the common assumption that larger datasets are always better for LLM training
• Results show practical improvements in test coverage and defect detection

For software engineering teams, this research offers a more efficient path to implementing AI-assisted testing tools by prioritizing data curation over massive data collection.

Less is More: On the Importance of Data Quality for Unit Test Generation