PaperlayerView exact revision on arXiv

Source-linked AI summary

On LLMs-Driven Synthetic Data Generation, Curation, and Evaluation: A Survey

Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, Haobo Wang

arXiv:2406.15126v1cs.CL

TL;DR

The paper addresses the lack of a unified framework for studying LLM-driven synthetic data generation amid persistent limitations in human-generated data. It surveys relevant work through a generic workflow spanning generation, curation, and evaluation, organizing the field, highlighting gaps, and outlining future research directions. The survey mainly focuses on text data and LLM-driven approaches while envisioning broader data-centric applications and LLM self-improvement.

  • Problem

    High-quality human data can be costly, scarce, privacy-sensitive, biased, and error-prone, while research on LLM-driven synthetic data lacks a unified framework.

  • Method

    The paper systematically surveys LLM-driven synthetic data generation, curation, and evaluation and proposes a generic workflow for real-world practice.

  • Results

    The survey organizes existing studies, identifies gaps and challenges, and proposes potential directions for future research.

  • Takeaways & Limitations

    The work offers guidance for building domain-specific datasets with LLMs and aims to support data production across various fields.

  • Takeaways & Limitations

    The survey mainly focuses on text data and LLM-driven approaches, leaving vision, speech, and other fields for future work.

Abstract

from arXiv · show

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.

1 Introduction

LLMs-driven synthetic data is presented as a scalable response to the cost, scarcity, privacy, bias, and error constraints of human-generated data. The survey organizes this emerging field around generation, curation, and evaluation to clarify its current gaps and future directions.

  • Human-generated data can be costly, scarce, privacy-sensitive, biased, and error-prone, motivating alternative collection methods.
  • LLMs can generate synthetic data that mimics real-world characteristics while providing knowledge-based fidelity and instruction-driven controllability.
  • LLMs-driven synthetic data can automate model training and evaluation with minimal human participation and provide scalable training and testing data.
  • Producing synthetic datasets with both high correctness and sufficient diversity requires careful process design, making the task non-trivial.
  • The survey organizes studies into generation, curation, and evaluation, identifies remaining gaps, and proposes directions for further development.

2 Preliminaries

The paper formalizes LLM-driven synthetic data generation as producing a dataset from a pretrained model, prompts, tasks, and optional supporting data. It emphasizes faithfulness and diversity as general requirements and frames generation, curation, and evaluation as coordinated components.

  • 2.1 Problem Definition: The generation task uses a pretrained LLM M to produce a final dataset Dgen from prompts p and a task specification T, optionally supported by Dsup.Dsup may consist of a small number of seed samples or unlabeled inputs.
  • 2.2 Requirements of Dgen: High-quality synthetic data should closely align with downstream evaluation metrics, although quality standards vary across tasks.
  • 2.2 Requirements of Dgen: Faithfulness requires logically and grammatically coherent data without hallucinations, factual errors, incorrect labels, or irrelevant content.These problems become more pronounced for long, complex, or domain-specific data.
  • 2.2 Requirements of Dgen: Diversity captures variation in length, topic, and writing style, helping synthetic data reflect real-world variation and prevent overfitting and bias.Uncontrolled LLM generation can instead become monotonous because of inherent model biases.
  • 2.2 Requirements of Dgen: Current approaches address faithfulness and diversity through three related topics: generation, curation, and evaluation.

3 Generic Workflow

The paper organizes LLM-driven synthetic data work into generation, curation, and evaluation, with generation commonly using prompt engineering or multi-step procedures. Prompt design controls task context, conditions, and demonstrations, while curation and evaluation address data quality and downstream usefulness.

  • LLM-driven synthetic data research generally covers generation, curation, and evaluation as interconnected parts of a generic workflow.
  • 3.1.1 Prompt Engineering: Prompt engineering guides generation through task specification, conditions, and in-context demonstrations wrapped in a natural-language template.Task specification provides context such as purpose, format, and background knowledge; demonstrations provide implicit guidance from real-world patterns.
  • 3.1.1 Prompt Engineering: Conditional prompting uses condition-value pairs to control dataset diversity, coverage, and focus when direct prompting produces repetitive outputs.Conditions may describe labels, topics, length, or style, with values sampled from known classes, retrieved from knowledge graphs, or generated by LLMs.
  • 3.1.1 Prompt Engineering: In-context examples can improve faithfulness, but their quality and diversity matter because prompt length and inconsistent demonstrations constrain effectiveness.Selecting examples focused on specific aspects is presented as an alternative to randomly sampling seed examples.
  • 3.1.2 Multi-Step Generation: Multi-step generation decomposes a complex task into scheduled subtasks whose intermediate outputs can condition later generations.The workflow includes sample-wise and dataset-wise decomposition, with the latter dynamically adjusting conditions to improve dataset coverage and direction.
  • 3.3 Data Evaluation: Direct evaluation assesses generated data itself, whereas indirect evaluation assesses its effectiveness on downstream tasks.For open-ended benchmarks without standardized answers, humans or auxiliary models use strategies such as rankings, ratings, or Elo scores.

4 Future Directions

The paper identifies future directions for making synthetic data generation more capable, knowledgeable, collaborative, and human-centered. These directions respond to limitations in reasoning, domain knowledge, model collaboration, and safe human involvement.

  • 4.1 Complex Task Decomposition: Complex real-world tasks expose limits in current multi-step generation, motivating agents that activate LLM reasoning and planning for autonomous data generation.The paper specifically proposes developing data-generation agents for industrial applications.
  • 4.2 Knowledge Enhancement: Knowledge-driven generation is proposed to address LLMs’ long-tailed and biased knowledge through external knowledge graphs or website retrieval.Enhanced domain knowledge may support feature definition, decomposition, reasoning, and generated-data evaluation.
  • 4.3 Synergy between Large & Small LMs: Future large-small model collaboration should move beyond prediction confidence toward diverse outputs and collaborative architectures.Real-time automated evaluation feedback is described as a way to guide adjustments during data generation.
  • Human involvement remains necessary because unconstrained synthetic data can contain noisy or toxic information and may poison models or cause mode collapse.The paper calls for human-centered interactive systems that consider feasibility, cost, readability, interpretability, and labor psychology.

5 Conclusion

The paper systematically reviews LLM-driven synthetic data generation and identifies challenges, opportunities, and future research directions.

  • The paper presents a systematic review of advancements in LLM-driven synthetic data generation.
  • It aims to guide enterprises and organizations in building domain-specific datasets with LLMs.
  • The review discusses challenges and opportunities while proposing potential directions for future research.
  • The authors envision LLM communities generating data for their own self-improvement.

Limitations

The survey focuses on text data and LLM-driven approaches, leaving synthetic-data research in other modalities for future work.

  • The survey mainly examines text data and LLM-driven approaches.
  • Synthetic data generation also spans modalities such as vision and speech, but these fields are outside the survey’s primary scope.
  • The authors plan to continue tracking newer work and add related approaches with more detailed analysis.

Ethics Statement

The proposed workflow may benefit data-centric researchers and industrial producers, while malicious uses of synthetic data raise ethical concerns.

  • The workflow can benefit researchers interested in data-centric AI and industrial producers facing data problems.
  • Malicious use of synthetic data raises ethical concerns requiring vigilance.

A Data Annotation

Selective annotation allocates samples between humans and LLMs, with uncertainty-based selection outperforming random selection in the cited results.

  • Selective annotation balances expensive, precise human annotation against more economical but rougher LLM-based annotation.
  • The key design problem is defining a cost-effective sample distribution between humans and LLMs.
  • Uncertainty-based methods, including maximal entropy and least confidence, outperform the random baseline.
  • These methods yield faster convergence and better downstream-model performance on annotated data than random selection.
  • Vote-k uses an unsupervised graph-based method to select diverse and representative examples for annotation.

B Tuning Techniques

The section surveys tuning techniques used to improve LLM-driven generation, focusing on model fine-tuning and soft prompting. These approaches have been extensively studied in other fields.

  • Model fine-tuning and soft prompting are the main tuning techniques highlighted for LLM-driven data generation.Both techniques are described as established subjects of study in other fields.
  • Existing tuning approaches are reported to improve generation performance.
  • The section directs readers to prior surveys and studies for detailed treatment of these techniques.

C Applications

LLM-driven synthetic data generation is presented as an alternative to human-dependent collection with applications spanning general, domain-specific, and multimodal tasks. These applications include standard NLP problems, costly specialized settings, and low-cost multimodal use cases.

  • LLM-driven generation serves as an alternative to traditional human-dependent data collection across general, domain-specific, and multimodal tasks.
  • Generic Tasks: Synthetic data generation is applied to general NLP tasks including classification, named entity recognition, question answering, relationship extraction, and natural language inference.
  • Domain-specific Tasks: Domain-specific applications include medical diagnosis, drug discovery, clinical trial extraction, industrial advertisement, and tabular data analysis.Human annotation can be extremely expensive and impractical in these settings.
  • Multimodal Tasks: Multimodal applications include text-image retrieval, chat understanding, visual question answering, and multimodal instruction tuning.The passage attributes this promise to the paradigm’s simplicity and low cost.

D Benchmark Datasets

The section summarizes benchmark datasets used to assess models trained with generated data. ToolBench is an LLM-generated benchmark for tool-use proficiency, while classification evaluations commonly test smaller models trained on generated data against existing benchmarks.

  • ToolBench is generated by LLMs and commonly evaluates LLM performance in tool usage proficiency.
  • In most classification evaluations, small language models trained on generated data are tested on existing benchmarks rather than using LLMs as test models.
  • Table 1 summarizes representative benchmark datasets for assessing models trained with generated data.
Loading 2406.15126v1…