Less is More: Adaptive Coverage for Synthetic Training Data

2025-04-20

AI Summary

In fast-response scenarios—such as emerging social trend identification and real-time network abuse detection—redundant synthetic data generated by large language models (e.g., Gemma, GPT) leads to inefficient training. Method: We propose an adaptive sampling method grounded in maximum coverage optimization—the first application of this combinatorial optimization paradigm to filtering LLM-generated synthetic training data. Leveraging a greedy approximation algorithm, our approach performs context-aware, dynamic subset selection. Results: On multi-class classification tasks, it significantly improves downstream classifier accuracy while reducing training data volume by 30–60%, accelerating fine-tuning and lowering computational cost. Our key contribution is establishing a “less-is-more” paradigm for synthetic data utilization: carefully curated, highly representative subsets consistently outperform full synthetic datasets—providing both theoretical grounding and practical tools for efficient, lightweight, synthetic-data-driven learning.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Synthetic training data generation with Large Language Models (LLMs) like Google's Gemma and OpenAI's GPT offer a promising solution to the challenge of obtaining large, labeled datasets for training classifiers. When rapid model deployment is critical, such as in classifying emerging social media trends or combating new forms of online abuse tied to current events, the ability to generate training data is invaluable. While prior research has examined the comparability of synthetic data to human-labeled data, this study introduces a novel sampling algorithm, based on the maximum coverage problem, to select a representative subset from a synthetically generated dataset. Our results demonstrate that training a classifier on this contextually sampled subset achieves superior performance compared to training on the entire dataset. This "less is more" approach not only improves accuracy but also reduces the volume of data required, leading to potentially more efficient model fine-tuning.

Problem

• Optimizing synthetic training data selection for classifiers
• Improving classifier accuracy with adaptive coverage sampling
• Reducing data volume for efficient model fine-tuning

Innovation

• Adaptive sampling algorithm for synthetic data
• Maximum coverage problem-based subset selection
• Improved accuracy with less training data