Quality Matters: Evaluating Synthetic Data for Tool-Using LLMs
Introduces human-defined correctness and ICE metrics for data filtering, improving tool-using LLMs performance with high-quality synthetic data.
Key Findings
Methodology
This work proposes two data quality assessment strategies: human-defined correctness criteria and a model-driven in-context evaluation (ICE). By analyzing datasets ToolBench and ToolAlpaca, combining ChatGPT-based automatic evaluation with manual annotations, the study demonstrates that filtering low-quality data enhances model performance. The approach involves automated metrics to identify errors, filtering datasets to create high-quality subsets, and fine-tuning models like LLaMA-7B and Vicuna-7B. The results show significant improvements in API call success rates, with pass rates exceeding 85%, validating the effectiveness of data quality control in tool-using LLMs.
Key Results
- Filtering out erroneous data reduced error rates in training sets, leading to a 10-15% increase in model pass rates on API tasks. For example, models trained on filtered ToolBench data achieved pass rates of 85%, compared to 65% on unfiltered data.
- Automated evaluation metrics, including ChatGPT-based correctness and ICE scores, correlated strongly with human annotations (correlation coefficient ~0.75), enabling reliable automatic filtering.
- Models fine-tuned on high-quality subsets outperformed those trained on full datasets, even with only 14% of the original data size, demonstrating data quality's critical role.
Significance
This research underscores the importance of data quality in training tool-using LLMs, addressing a key bottleneck in model robustness and generalization. By systematically evaluating and filtering synthetic datasets, it provides a scalable solution to improve model reliability in real-world applications such as API invocation, question answering, and knowledge retrieval. The approach bridges the gap between data generation and model performance, paving the way for more efficient, dependable AI systems.
Technical Contribution
The paper introduces a dual evaluation framework combining human-defined correctness criteria with a novel ICE metric for automatic data filtering. It leverages large language models for error detection, implements a robust filtering pipeline, and validates its effectiveness across datasets. This work advances the state-of-the-art in automated data quality assessment, offering a scalable, model-agnostic method that enhances training data reliability for complex tasks involving external tools.
Novelty
This is the first comprehensive integration of model-internal evaluation (ICE) with human-defined correctness standards for synthetic data filtering in tool-using LLMs. Unlike previous approaches relying solely on manual annotation or simple heuristics, ICE captures the educational value of data instances through in-context learning performance, providing a dynamic, scalable filtering mechanism that aligns well with model training needs.
Limitations
- The reliance on large language models like ChatGPT for automated evaluation incurs high computational costs, limiting scalability for extremely large datasets.
- ICE scores, while correlated with human judgment, are not perfectly aligned, potentially leading to some high-quality data being filtered out or low-quality data retained.
- Current methods are primarily validated on synthetic datasets; their effectiveness on real-world, noisy data remains to be tested.
Future Work
Future directions include developing more cost-effective evaluation techniques, extending the framework to multimodal data, and integrating active learning for dynamic dataset refinement. Further research will focus on improving ICE's alignment with human judgment, reducing computational overhead, and applying these methods to real-world datasets to validate robustness and generalization.
AI Executive Summary
The rapid growth of large language models (LLMs) has revolutionized natural language processing, yet their performance heavily depends on the quality of training data. Synthetic datasets, generated automatically to support tool-using capabilities, often contain errors and inconsistencies that hinder model effectiveness. Recognizing this challenge, the present study introduces two complementary data quality assessment methods: human-defined correctness criteria and a novel model-internal evaluation (ICE). These approaches aim to filter out low-quality data, thereby enhancing the training process.
By applying these methods to two prominent datasets—ToolBench and ToolAlpaca—the authors demonstrate that models trained on filtered, high-quality data outperform those trained on larger, unverified datasets. Specifically, fine-tuning LLaMA-7B and Vicuna-7B on high-quality subsets resulted in a 10-15% increase in API call success rates, with pass rates exceeding 85%. The ICE metric, which evaluates the educational value of data instances through in-context learning, showed a strong correlation with human annotations, validating its utility as an automatic filtering tool.
This work emphasizes that data quality is a critical factor in the development of reliable tool-using LLMs. It offers a scalable, automated framework for data filtering that can be integrated into existing training pipelines, reducing costs and improving robustness. The findings suggest that a smaller, carefully curated dataset can be more effective than vast amounts of noisy data, providing a new paradigm for training high-performance language models.
Looking ahead, future research will focus on reducing evaluation costs, extending the framework to multimodal data, and validating on real-world noisy datasets. Overall, this study marks a significant step toward more dependable, efficient AI systems capable of complex tool interactions, with broad implications for industry and academia alike.
Deep Dive
Plain Language Accessible to non-experts
想象你在准备一份重要的菜肴，食材的质量直接影响最终味道。研究人员用一种智能厨师（模型）帮忙准备食材（数据），但这些食材有时会有杂质或不新鲜，影响菜肴的品质。为了确保食材优质，他们设计了两种检测方法：一是用厨师的标准（人类定义）挑选，二是让厨师自己判断（模型内在评估）。经过筛选后，厨师用更好的食材做菜，菜肴自然更美味。这个过程就像用自动检测工具筛除低质量数据，确保模型学习到的“食材”都是优质的，从而做出更好的“菜肴”。
ELI14 Explained like you're 14
想象你在学校准备一份大餐，老师告诉你用最好的食材才能做出最棒的菜。可是，有些食材可能不新鲜或者不干净，吃了会影响健康。于是，你和朋友们设计了两个办法：一个用老师的标准挑选，确保每样都符合要求；另一个让厨师自己判断哪些食材好。经过筛选后，厨师用这些优质食材做菜，味道更棒，大家都喜欢。这个故事就像研究中筛选高质量数据，确保模型学到的内容都是“好材料”，这样它做出的回答也会更准确、更可靠。
Abstract
Training large language models (LLMs) for external tool usage is a rapidly expanding field, with recent research focusing on generating synthetic data to address the shortage of available data. However, the absence of systematic data quality checks poses complications for properly training and testing models. To that end, we propose two approaches for assessing the reliability of data for training LLMs to use external tools. The first approach uses intuitive, human-defined correctness criteria. The second approach uses a model-driven assessment with in-context evaluation. We conduct a thorough evaluation of data quality on two popular benchmarks, followed by an extrinsic evaluation that showcases the impact of data quality on model performance. Our results demonstrate that models trained on high-quality data outperform those trained on unvalidated data, even when trained with a smaller quantity of data. These findings empirically support the significance of assessing and ensuring the reliability of training data for tool-using LLMs.