Cost-Effective Synthetic Data Generation for Post-Training using QWICK
Keywords: synthetic data generation, post-training
Abstract: Large language models (LLMs) are showing expert-level ability in various fields (e.g., programming and math). However, this progress heavily relies on the generation of high-quality synthetic data to improve the models’ capabilities during post-training. Generating such data in a cost-effective manner presents a significant challenge. Specifically, stronger models tend to generate higher-quality data but come with a substantial computational cost, while weaker models are cheaper to run but may produce weaker outputs. In this paper, we introduce Question-Wise model pICK (QWICK) to address this challenge. By tracking the empirical reward, cost, and number of trials for each model, QWICK strikes a balance between exploitation and exploration, ultimately converging on a cost-effective model for each specific question. Specifically, QWICK achieves a 50\% cost reduction on a programming dataset and a 40\% cost reduction on a mathematics dataset, without compromising data quality. Furthermore, compared to baseline methods, our approach can produce up to 2.1 times more valid synthetic data at the same cost. Our anonymized code is available at https://anonymous.4open.science/r/QWICK-17C3
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2257
Loading