Project Structure

ChildEval/
├── prompts/                # prompts for data generation
├── data/                   # full cleaned data and partial raw generation examples
├── evaluation_scripts/      # evaluation scripts and corresponding prompts
└── README.md               # project description

Note

In the original persona construction, we generated nicknames for Chinese children. Although these nicknames were synthetically created, we removed and anonymized them prior to data release to mitigate any potential privacy concerns.

All released data has been manually reviewed to ensure content safety. The dataset is primarily generated for Chinese children and all English data is obtained via translation using Qwen2.5-72B-instruct. In the data folder, we provide a subset of generated examples, where the keys task and task_dialog correspond to explicit preferences and implicit preferences, respectively.

All examples in the dataset are fully generated by models without manual rewriting; human involvement is limited to safety verification to ensure that no harmful or inappropriate content is included.

Disclaimer

All data in this dataset is synthetically generated. Although we have conducted manual checks to ensure safety and appropriateness, we cannot guarantee that the dataset is entirely free from unintended biases, inaccuracies, or potentially sensitive content. The dataset is provided “as is” for research purposes only, and users are responsible for conducting their own validation and risk assessment before deploying it in real-world applications.