The "Can't Talk About This" topic control dataset, released by NVIDIA in 2024, focuses on dialogue topic management and safe alignment of large language models; related research findings can be found in the paper titled «CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues”, aimed at improving large language models’ ability to maintain topic focus in task-oriented dialogues and enhancing their robustness against distractions.

The dataset contains 1,080 synthetic dialogue samples covering nine domains: health, banking, travel, education, finance, insurance, law, real estate, and computer troubleshooting. Each dialogue includes distracting turns designed to test the model’s resistance to topic drift. Fine-tuning on this dataset significantly enhances performance in instruction-following and safety-related tasks, enabling effective identification of sensitive topics and handling of restricted content.

Dataset Composition

The dataset primarily consists of the following fields:

• domain: The field or category the dialogue belongs to
• scenario: The specific context or task being discussed
• system_instruction: Dialogue strategy instructions provided to the model, typically containing complex sets of rules specifying allowed and prohibited discussion topics
• conversation: Full dialogue content, including main-topic exchanges and distracting turns
• distractors: List of distracting turns, comprising bot responses along with corresponding user replies intended as counter-responses to those bot turns
• conversation_with_distractors: Complete dialogue incorporating all distracting elements

The dataset is split into training and testing subsets. The training set was synthetically generated using the GPT-4 Turbo model, while the test set comprises human-labeled evaluation data featuring more complex and realistic distractor examples for assessing model performance.

Citation```bibtex

@inproceedings{sreedhar2024canttalkaboutthis, title={CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues}, author={Sreedhar, Makesh and Rebedea, Traian and Ghosh, Shaona and Zeng, Jiaqi and Parisien, Christopher}, booktitle={Findings of the Association for Computational Linguistics: EMNLP 2024}, pages={12232--12252}, year={2024}, organization={Association for Computational Linguistics} }