Reproduce “Balancing Cost and Effectiveness of Synthetic Data Generation Strategies for LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yung-Chieh Chan, George Pu, Apaar Shanker, Parth Suresh, Penn Jenks, John Heyer, Sam Denton
Researchers from Scale AI systematically evaluate various synthetic data generation strategies for fine-tuning Large Language Models (LLMs), determining that the optimal approach depends on the ratio of the teacher model query budget to the seed instruction set size. The study shows "Answer Augmentation" is effective for low budget ratios, while "New Question Evolution" performs better at high budget ratios.
Key insights
- The choice of the optimal synthetic data generation strategy is highly dependent on the Budget Ratio (query budget relative to seed instruction set size).
- The capability of the augmenter model (πaug) significantly impacts the performance of "New Question Evolution" but has little effect on "Question Rephrasing."
- Verifying synthetic responses against ground truth answers to filter out incorrect data did not yield significant improvements in student model accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.