[Paper Review] On LLMs-Driven Synthetic Data Generation, Curation, and Evaluation: A Survey

Lin Long, Rui Wang+5 more|arXiv (Cornell University)|Jun 14, 2024

Scientific Computing and Data Management4 citations

TL;DR

This survey presents a unified framework for LLM-driven synthetic data generation, curation, and evaluation, organizing existing methods into a coherent workflow. It identifies critical gaps in data quality, diversity, and evaluation rigor, and advocates for systematic, methodical approaches to advance both research and industrial applications in data-centric AI.

ABSTRACT

Within the evolving landscape of deep learning, the dilemma of data quantity and quality has been a long-standing problem. The recent advent of Large Language Models (LLMs) offers a data-centric solution to alleviate the limitations of real-world data with synthetic data generation. However, current investigations into this field lack a unified framework and mostly stay on the surface. Therefore, this paper provides an organization of relevant studies based on a generic workflow of synthetic data generation. By doing so, we highlight the gaps within existing research and outline prospective avenues for future study. This work aims to shepherd the academic and industrial communities towards deeper, more methodical inquiries into the capabilities and applications of LLMs-driven synthetic data generation.

Motivation & Objective

• To address the lack of a unified framework in LLM-driven synthetic data research, which currently remains fragmented and superficial.
• To systematize the current landscape by organizing studies into a coherent workflow of generation, curation, and evaluation.
• To highlight critical gaps in data quality, diversity, and evaluation practices that hinder robust model training and deployment.
• To guide academic and industrial communities toward more methodical and effective use of synthetic data in NLP and beyond.
• To promote the use of synthetic data as a scalable, controllable, and high-quality alternative to human-annotated data in deep learning.

Proposed method

• Proposes a generic, three-stage workflow: synthetic data generation, curation, and evaluation, serving as a unifying structure for existing research.
• Categorizes and analyzes existing methods based on their approach to prompt engineering, few-shot prompting, chain-of-thought prompting, and self-consistency techniques.
• Introduces a taxonomy of data curation strategies, including filtering, distillation, and active learning, to improve data quality and reduce bias.
• Reviews evaluation protocols that assess both fluency and task-specific performance, including zero-shot and few-shot generalization on benchmark datasets.
• Emphasizes the role of instruction-following capabilities in enabling controllable data generation for specific downstream tasks.
• Analyzes the use of LLMs not only as generators but also as evaluators and curators, enabling end-to-end synthetic data pipelines with minimal human intervention.

Experimental results

Research questions

• RQ1 What are the core components and stages of a systematic LLM-driven synthetic data pipeline, and how can they be unified into a coherent framework?
• RQ2 How do existing methods ensure high correctness and sufficient diversity in synthetic data across different NLP tasks and domains?
• RQ3 What are the key limitations in current evaluation practices for synthetic data, and how can they be improved to reflect real-world model performance?
• RQ4 In what ways can LLMs be leveraged not only for generation but also for curation and evaluation of synthetic datasets?
• RQ5 What are the most promising applications and future research directions for synthetic data in data-centric AI, particularly in low-resource or high-bias scenarios?

Key findings

• Over 300 datasets on Hugging Face are already tagged as 'synthetic', indicating widespread adoption and integration into mainstream LLM training pipelines.
• LLM-generated datasets such as Alpaca, Vicuna, OpenHermes 2.5, and OpenChat 3.5 have become foundational for fine-tuning and pre-training, demonstrating real-world viability.
• Despite progress, many approaches still rely on accessible LLMs, with limited exploration of their use with black-box models, indicating a key research gap.
• Synthetic data generation is most effective when combined with curation techniques like filtering and distillation, which significantly improve data quality and reduce bias.
• Evaluation benchmarks like ToolBench, TruthfulQA, and MATH show that models fine-tuned on synthetic data can achieve strong performance, especially when data quality and diversity are carefully managed.
• The study identifies a critical need for standardized evaluation protocols and better alignment between synthetic data characteristics and downstream task requirements to ensure robustness and fairness.