Unmasking and Improving Data Credibility: A Study with Datasets for Training Harmless Language Models

TL;DR

This paper presents a systematic framework and an open-source tool for evaluating and improving the credibility of datasets used in training harmless language models, significantly reducing label errors and enhancing model safety.

Contribution

The study introduces a novel framework for dataset credibility assessment and demonstrates its effectiveness by fixing label errors in multiple benchmarks, improving data quality for safe language model training.

Findings

Average of 6.16% label errors found and fixed in datasets

Data cleaning improves downstream model performance

Open-source tool Docta facilitates dataset cleaning

Abstract

Language models have shown promise in various tasks but can be affected by undesired data during training, fine-tuning, or alignment. For example, if some unsafe conversations are wrongly annotated as safe ones, the model fine-tuned on these samples may be harmful. Therefore, the correctness of annotations, i.e., the credibility of the dataset, is important. This study focuses on the credibility of real-world datasets, including the popular benchmarks Jigsaw Civil Comments, Anthropic Harmless & Red Team, PKU BeaverTails & SafeRLHF, that can be used for training a harmless language model. Given the cost and difficulty of cleaning these datasets by humans, we introduce a systematic framework for evaluating the credibility of datasets, identifying label errors, and evaluating the influence of noisy labels in the curated language data, specifically focusing on unsafe comments and conversation classification. With the framework, we find and fix an average of 6.16% label errors in 11 datasets constructed from the above benchmarks. The data credibility and downstream learning performance can be remarkably improved by directly fixing label errors, indicating the significance of cleaning existing real-world datasets. We provide an open-source tool, Docta, for data cleaning at https://github.com/Docta-ai/docta.