Demystifying Synthetic Data in LLM Pre-training: A Systematic Study of Scaling Laws, Benefits, and Pitfalls

• URL: http://arxiv.org/abs/2510.01631v1
• Date: Thu, 02 Oct 2025 03:24:42 GMT
• Title: Demystifying Synthetic Data in LLM Pre-training: A Systematic Study of Scaling Laws, Benefits, and Pitfalls
• Authors: Feiyang Kang, Newsha Ardalani, Michael Kuchnik, Youssef Emad, Mostafa Elhoushi, Shubhabrata Sengupta, Shang-Wen Li, Ramya Raghavendra, Ruoxi Jia, Carole-Jean Wu,
• Abstract summary: Training data plays a crucial role in Large Language Models (LLM) scaling, yet high quality data is of limited supply.<n>We compare natural web data, diverse synthetic types (rephrased text, generated textbooks), and mixtures of natural and synthetic data.<n>We find pre-training on rephrased synthetic data textitalone is not faster than pre-training on natural web texts.
• Score: 25.294408301653576
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Training data plays a crucial role in Large Language Models (LLM) scaling, yet high quality data is of limited supply. Synthetic data techniques offer a potential path toward sidestepping these limitations. We conduct a large-scale empirical investigation (>1000 LLMs with >100k GPU hours) using a unified protocol and scaling laws, comparing natural web data, diverse synthetic types (rephrased text, generated textbooks), and mixtures of natural and synthetic data. Specifically, we found pre-training on rephrased synthetic data \textit{alone} is not faster than pre-training on natural web texts; while pre-training on 1/3 rephrased synthetic data mixed with 2/3 natural web texts can speed up 5-10x (to reach the same validation loss) at larger data budgets. Pre-training on textbook-style synthetic data \textit{alone} results in notably higher loss on many downstream domains especially at small data budgets. "Good" ratios of synthetic data in training data mixtures depend on the model size and data budget, empirically converging to ~30% for rephrased synthetic data. Larger generator models do not necessarily yield better pre-training data than ~8B-param models. These results contribute mixed evidence on "model collapse" during large-scale single-round (n=1) model training on synthetic data--training on rephrased synthetic data shows no degradation in performance in foreseeable scales whereas training on mixtures of textbook-style pure-generated synthetic data shows patterns predicted by "model collapse". Our work demystifies synthetic data in pre-training, validates its conditional benefits, and offers practical guidance.

Related papers

• Contrastive Decoding for Synthetic Data Generation in Low-Resource Language Modeling [9.380879437204277]
  We investigate the benefits of contrastive decoding for generating synthetic corpora.<n>By amplifying the signal from a model that has better performance, we create a synthetic corpus and mix it with the original training data.<n>Our findings show that training on a mixture of synthesized and real data improves performance on the language modeling objective and a range of downstream tasks.
  arXiv  Detail & Related papers  (2025-10-09T14:04:52Z)
• Understanding the Influence of Synthetic Data for Text Embedders [52.04771455432998]
  We first reproduce and publicly release the synthetic data proposed by Wang et al.<n>We critically examine where exactly synthetic data improves model generalization.<n>Our findings highlight the limitations of current synthetic data approaches for building general-purpose embedders.
  arXiv  Detail & Related papers  (2025-09-07T19:28:52Z)
• Less is More: Adaptive Coverage for Synthetic Training Data [20.136698279893857]
  This study introduces a novel sampling algorithm, based on the maximum coverage problem, to select a representative subset from a synthetically generated dataset.<n>Our results demonstrate that training a classifier on this contextually sampled subset achieves superior performance compared to training on the entire dataset.
  arXiv  Detail & Related papers  (2025-04-20T06:45:16Z)
• Scaling Laws of Synthetic Data for Language Models [132.67350443447611]
  We introduce SynthLLM, a scalable framework that transforms pre-training corpora into diverse, high-quality synthetic datasets.<n>Our approach achieves this by automatically extracting and recombining high-level concepts across multiple documents using a graph algorithm.
  arXiv  Detail & Related papers  (2025-03-25T11:07:12Z)
• How to Synthesize Text Data without Model Collapse? [37.219627817995054]
  Model collapse in synthetic data indicates that iterative training on self-generated data leads to a gradual decline in performance.<n>We propose token editing on human-produced data to obtain semi-synthetic data.
  arXiv  Detail & Related papers  (2024-12-19T09:43:39Z)
• Little Giants: Synthesizing High-Quality Embedding Data at Scale [71.352883755806]
  We introduce SPEED, a framework that aligns open-source small models to efficiently generate large-scale embedding data. SPEED uses only less than 1/10 of the GPT API calls, outperforming the state-of-the-art embedding model E5_mistral when both are trained solely on their synthetic data.
  arXiv  Detail & Related papers  (2024-10-24T10:47:30Z)
• Online Data Augmentation for Forecasting with Deep Learning [0.33554367023486936]
  This work introduces an online data augmentation framework that generates synthetic samples during the training of neural networks.<n>We maintain a balanced representation between real and synthetic data throughout the training process.<n>Experiments suggest that online data augmentation leads to better forecasting performance compared to offline data augmentation or no augmentation approaches.
  arXiv  Detail & Related papers  (2024-04-25T17:16:13Z)
• Let's Synthesize Step by Step: Iterative Dataset Synthesis with Large Language Models by Extrapolating Errors from Small Models [69.76066070227452]
  *Data Synthesis* is a promising way to train a small model with very little labeled data. We propose *Synthesis Step by Step* (**S3**), a data synthesis framework that shrinks this distribution gap. Our approach improves the performance of a small model by reducing the gap between the synthetic dataset and the real data.
  arXiv  Detail & Related papers  (2023-10-20T17:14:25Z)
• On the Stability of Iterative Retraining of Generative Models on their own Data [56.153542044045224]
  We study the impact of training generative models on mixed datasets. We first prove the stability of iterative training under the condition that the initial generative models approximate the data distribution well enough. We empirically validate our theory on both synthetic and natural images by iteratively training normalizing flows and state-of-the-art diffusion models.
  arXiv  Detail & Related papers  (2023-09-30T16:41:04Z)
• From Fake to Real: Pretraining on Balanced Synthetic Images to Prevent Spurious Correlations in Image Recognition [64.59093444558549]
  We propose a simple, easy-to-implement, two-step training pipeline that we call From Fake to Real. By training on real and synthetic data separately, FFR does not expose the model to the statistical differences between real and synthetic data. Our experiments show that FFR improves worst group accuracy over the state-of-the-art by up to 20% over three datasets.
  arXiv  Detail & Related papers  (2023-08-08T19:52:28Z)
• Synthetic data, real errors: how (not) to publish and use synthetic data [86.65594304109567]
  We show how the generative process affects the downstream ML task. We introduce Deep Generative Ensemble (DGE) to approximate the posterior distribution over the generative process model parameters.
  arXiv  Detail & Related papers  (2023-05-16T07:30:29Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.