Donkii: Can Annotation Error Detection Methods Find Errors in Instruction-Tuning Datasets?
This work addresses the challenge of data quality in instruction-tuning for LLMs, which is crucial for improving model performance, though it is incremental as it extends AED methods from classification to generation tasks.
The paper tackles the problem of applying Annotation Error Detection (AED) methods to instruction-tuning datasets for Large Language Models, finding that existing datasets contain errors that propagate into models, and proposes four AED baselines with practical recommendations for cleaning such data.
Instruction tuning has become an integral part of training pipelines for Large Language Models (LLMs) and has been shown to yield strong performance gains. In an orthogonal line of research, Annotation Error Detection (AED) has emerged as a tool for detecting quality problems in gold standard labels. So far, however, the application of AED methods has been limited to classification tasks. It is an open question how well AED methods generalize to language generation settings, which are becoming more widespread via LLMs. In this paper, we present a first and novel benchmark for AED on instruction tuning data: DONKII. It comprises three instruction-tuning datasets enriched with error annotations by experts and semi-automatic methods. We also provide a novel taxonomy of error types for instruction-tuning data. We find that all three datasets contain clear errors, which sometimes propagate directly into instruction-tuned LLMs. We propose four AED baselines for the generative setting and evaluate them extensively on the newly introduced dataset. Our results show that the choice of the right AED method and model size is indeed crucial and derive practical recommendations for how to use AED methods to clean instruction-tuning data.