CantTalkAboutThis Topic Control Dataset - Non Commercial is a dataset released by NVIDIA in 2024, focusing on conversational topic control and safety alignment of language models. The associated paper is CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues, aiming to enhance the ability of large language models to maintain topic focus in task-oriented dialogues and improve their robustness against distractors.

The dataset contains 1,080 synthetic dialogue samples across nine domains: health, banking, travel, education, finance, insurance, law, real estate, and computer troubleshooting. Each dialogue includes distractor turns designed to test the model's resistance to topic deviation.

Fine-tuning on this dataset can significantly improve model performance in instruction-following and safety tasks, enabling effective identification of sensitive topics and handling of restricted content.

Dataset Composition

The dataset primarily includes the following fields:

• domain: The domain to which the conversation belongs.
• scenario: The specific scenario or task being discussed.
• system_instruction: The conversational strategy assigned to the model, typically containing a complex set of instructions specifying allowed and prohibited discussion topics.
• conversation: The complete dialogue content, including the main topic and distractor turns.
• distractors: A list of distractor turns, comprising bot turns within the dialogue and user responses intended as reactions to those bot turns.
• conversation_with_distractors: The complete dialogue including the distractor turns.

The dataset is divided into a training set and a test set. The training set was synthetically generated by the GPT-4 Turbo model, while the test set is a human-annotated evaluation dataset containing more complex and realistic distractors for assessing model performance.

Citation

@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}
}