Evaluating Language Models as Synthetic Data Generators (2412.03679v1)

Abstract: Given the increasing use of synthetic data in LLM (LM) post-training, an LM's ability to generate high-quality data has become nearly as crucial as its ability to solve problems directly. While prior works have focused on developing effective data generation methods, they lack systematic comparison of different LMs as data generators in a unified setting. To address this gap, we propose AgoraBench, a benchmark that provides standardized settings and metrics to evaluate LMs' data generation abilities. Through synthesizing 1.26 million training instances using 6 LMs and training 99 student models, we uncover key insights about LMs' data generation capabilities. First, we observe that LMs exhibit distinct strengths. For instance, GPT-4o excels at generating new problems, while Claude-3.5-Sonnet performs better at enhancing existing ones. Furthermore, our analysis reveals that an LM's data generation ability doesn't necessarily correlate with its problem-solving ability. Instead, multiple intrinsic features of data quality-including response quality, perplexity, and instruction difficulty-collectively serve as better indicators. Finally, we demonstrate that strategic choices in output format and cost-conscious model selection significantly impact data generation effectiveness.

An Analytical Overview of "Evaluating LLMs as Synthetic Data Generators"

The paper "Evaluating LLMs as Synthetic Data Generators" systematically addresses a pressing issue in the NLP community: assessing the capabilities of LLMs (LMs) in generating high-quality synthetic data. This domain is critical as generating synthetic data is increasingly recognized as a scalable complement to manual data annotation, which can enhance model performance over various tasks. The authors introduce a novel benchmark, AGORA BENCH, designed to provide a standardized framework for evaluating the data generation capabilities of LMs.

Core Contributions and Methodology

The methodology involves synthesizing 1.26 million training instances using six different LMs and training 99 student models to derive insights into the LMs' synthetic data-generation capabilities. Key experimental domains include mathematics, instruction-following, and code, with data generation methods categorized into instance generation, response generation, and quality enhancement.

AGORA BENCH provides a rigorous framework where critical variables such as meta-prompts and seed datasets remain constant, ensuring that the variability in performance is attributable solely to the LMs' data generation capabilities.

Key Insights and Findings

One of the paper's pivotal observations is the differentiation in strengths among various LMs. Notably, GPT-4o excels in generating novel instances, while Claude-3.5-Sonnet performs better in quality enhancement tasks. The paper reveals an unexpected trend; LMs with weaker problem-solving capabilities occasionally surpass more robust counterparts in data generation efficiency. This phenomenon suggests that the data generation ability of an LM cannot be solely predicted based on its problem-solving prowess.

Moreover, several intrinsic features, including instruction difficulty, response quality, and response perplexity, collectively show a correlation with the effectiveness of the student model's improvement, rather than the LM's inherent problem-solving capacity.

Theoretical and Practical Implications

The findings underscore the importance of strategic LM selection tailored to specific data generation needs. Additionally, the interpretation of intrinsic data features can serve as predictive indicators for data generation efficiency, offering a new dimension to LM evaluation beyond conventional benchmarks.

Practically, these insights encourage the community to consider cost-effective strategies in LM deployment, such as generating larger volumes of data with cheaper models, which sometimes yields better results than using fewer data points from expensive models.

Directions for Future Research

As the field progresses, this benchmark will likely catalyze two significant advancements: the development of LMs precisely tuned for data generation and the refinement of evaluation frameworks across diverse NLP pipelines. The benchmark's extensibility allows for integration with custom data scenarios, enhancing its utility for both academic and commercial applications.

The paper sets a foundational precedent for comprehensive LM assessment protocols, encouraging further exploration into how intrinsic qualities of synthetic data can be optimized to aid the crafting of more competent, contextually relevant, and adaptable NLP applications.