View a PDF of the paper titled Something Just Like TRuST: Toxicity Recognition of Span and Target, by Berk Atil and two other authors.
View PDF | HTML (experimental)
Abstract: Toxic language includes content that is offensive, abusive, or that promotes harm. Progress in preventing toxic output from large language models (LLMs) is hampered by inconsistent definitions of toxicity. We introduce TRuST, a large-scale dataset that unifies and expands prior resources through a carefully synthesized definition of toxicity and corresponding annotation scheme. It consists of ~300k annotations, with high-quality human annotation on ~11k. To ensure high quality, we designed a rigorous, multi-stage human annotation process and evaluated the diversity of the annotators. Then we benchmarked state-of-the-art LLMs and pre-trained models on three tasks: toxicity detection, identification of the target group, and of toxic words. Our results indicate that fine-tuned PLMs outperform LLMs on the three tasks, and that current reasoning models do not reliably improve performance. TRuST constitutes one of the most comprehensive resources for evaluating and mitigating LLM toxicity and other research in socially-aware and safer language technologies.
Submission History
From: Berk Atil [view email]
[v1] Mon, 2 Jun 2025 23:48:16 UTC (1,094 KB)
[v2] Mon, 5 Jan 2026 21:38:57 UTC (1,098 KB)
—
### Understanding Toxicity in Language Models
Toxic language can take many forms, often presenting as offensive, abusive, or harmful content. In today’s digital landscape, large language models (LLMs) are becoming increasingly integrated into everyday applications, from customer service to content creation. However, the challenge of ensuring these models do not propagate toxic language remains a pressing issue. This calls for a comprehensive understanding of toxicity and how to effectively manage it within LLMs.
### The TRuST Dataset: A New Benchmark
The introduction of the TRuST dataset marks a significant advancement in the field of toxicity recognition. This large-scale dataset consolidates and expands upon previous resources, creating a comprehensive framework for defining toxicity. With approximately 300,000 annotations, of which around 11,000 have undergone high-quality human annotation, TRuST aims to provide a robust foundation for future research. The dataset is instrumental for developers and researchers seeking to refine the capabilities of LLMs by providing a clear understanding of what constitutes toxic language.
More Read
### Multi-Stage Annotation Process
The quality of any dataset is paramount, and TRuST has been developed through a meticulous multi-stage human annotation process. This was designed to not only ensure the accuracy of the toxicity classifications but also to evaluate the diversity of the annotators. Diversity among annotators helps to reduce bias in the dataset, ensuring it reflects a variety of perspectives and cultural contexts. This thoughtful approach to annotation signifies a leap forward in addressing the complexities surrounding toxic language.
### Benchmarking LLMs and Pre-trained Models
The efficacy of any dataset can be evaluated through benchmarking against existing models. In the case of TRuST, state-of-the-art LLMs and pre-trained models were assessed on three primary tasks: toxicity detection, identification of the target group, and pinpointing toxic words. The findings revealed that fine-tuned pre-trained language models (PLMs) significantly outperform LLMs in these tasks. This insight is crucial for developers aiming to build safer language technologies, as it informs the selection of models based on specific functionalities.
### Addressing Current Limitations
One of the key discoveries in the TRuST study is that the current reasoning models do not consistently enhance performance in toxicity detection. This indicates a critical area for future research, highlighting the need for ongoing development to improve the capabilities of these models. It brings to light the importance of understanding the limitations of current technologies and the necessity for continual innovation in the field of natural language processing.
### Implications for Safer Language Technologies
TRuST stands out as one of the most comprehensive resources available for evaluating and mitigating toxicity in large language models. It paves the way for further research into socially-aware language technologies, enabling developers and researchers to create applications that are not only effective but also responsible. By utilizing the TRuST dataset, stakeholders can contribute to building a digital landscape that prioritizes safety, inclusivity, and respect, making strides towards a more harmonious online environment.
### Future Directions in Toxicity Research
As the discourse surrounding toxic language continues to evolve, the need for innovative solutions remains significant. Future research should focus on enhancing the methodologies used to identify and mitigate toxicity, as well as expanding datasets like TRuST to encapsulate a broader spectrum of language nuances. By fostering collaboration within the research community, we can collectively work towards developing more sophisticated tools that effectively curb the spread of toxic language in digital communication.
—
This article serves as a detailed exploration of the TRuST dataset and its implications in the fight against toxic language in AI. By focusing on critical aspects such as annotation quality, benchmarking, and future directions for research, it provides a well-rounded understanding of the ongoing challenges and advancements in this essential area of study.
Inspired by: Source
