Online filtering improves synthetic data training with real data guidance
Let Training Guide Selection: Online Synthetic Data Filtering via Real-Anchored Utility
Machine Learning
Summary
Training machine learning models with fake (synthetic) data can help when there’s not enough real data, but sometimes the fake data isn’t very useful or accurate. The authors developed a method called FROST that decides which synthetic data to keep by checking how much it helps improve the model based on real data examples. This method filters out some low-value synthetic data during training without needing extra test sets, making training more efficient and often more accurate. They tested it in multiple tasks including image recognition, language understanding, and advertising, and it improved performance while using less synthetic data.
What this means in practice
- •For machine learning engineers: Improve model training by automatically filtering synthetic data batches based on utility feedback from real data without extra validation sets.
- •For advertising technology teams: Achieve better performance in ad re-ranking systems by applying real-anchored filtering of synthetic training data during model updates.$Commercial implications: Enables a product feature that optimizes ad ranking models using fewer but more effective synthetic examples, improving ad relevance.
Authors
Yanran Wu, Sana Lakdawala, Renzo Tassara Miller, Chongyang Bai, Sharath Ciddu, Shivendra Pratap Singh, Kungang Li, Sandeep Pandey, Chunwei Liu
Abstract
Synthetic data can scale training supervision when real-world data are limited, but noise and distribution mismatch can reduce its value. Existing synthetic data selection methods often emphasize fidelity or diversity rather than the learner's evolving needs. We propose FROST, an online framework that estimates synthetic-data utility through gradient feedback anchored in real training data. It calibrates batch utility against recent history to determine when filtering is needed and filters samples only in out-of-band batches to determine what to retain, without an external verifier or held-out validation set. Experiments on two public benchmarks for image classification and LLM fine-tuning for text-to-SQL show that FROST filters out around 20--30% of the synthetic data while improving real-task performance compared with training on the full synthetic data pool. We further apply FROST during training in a large-scale industrial ads re-ranking system, achieving significant performance gains over a highly optimized production baseline, demonstrating its effectiveness and generalizability.