ChildSafeLLM: A Dataset of Child Safety Aligned Evaluation Prompts for Generative AIs

ChildSafeLLM (Child Safety Aligned Evaluation Prompts for Generative AIs) is a developmental benchmark comprising 200 total prompts carefully crafted and curated for two distinct age cohorts—ages 6-12 and ages 7-13. Drawing on evidence-based child-safety guidelines, school curricula, and authentic questions from children and caregivers, the prompts were sourced from reputable organizations across six continents to ensure broad cultural and linguistic representation. Available in .xlsx, ChildSafeLLM supports a wide range of natural-language processing tasks, including harmful-content detection, age-appropriateness classification, response-consistency analysis, safety-alignment fine-tuning, and developmental language modeling. Beyond its use for model training, the corpus offers a reproducible benchmark for evaluating ambiguity resolution, risk-category labeling, and multi-turn dialogue safety in child–AI interactions. By standardizing safety scenarios for younger users, ChildSafeLLM advances research into trustworthy AI design and helps practitioners assess how generative models accommodate children’s cognitive and emotional needs. (2025-06-10)面向生成式AI（Generative AI）的儿童安全对齐评估提示集（ChildSafeLLM，Child Safety Aligned Evaluation Prompts for Generative AIs）是一款儿童发展基准测试集，总计包含200条精心打磨与甄选的提示词，专为6-12岁及7-13岁两个独立年龄群体设计。该提示集依托循证儿童安全指南、学校课程体系，以及儿童与看护者的真实提问构建，其来源覆盖六大洲的权威机构，以确保具备广泛的文化与语言代表性。该数据集以.xlsx格式发布，支持涵盖多项自然语言处理（Natural Language Processing, NLP）任务，包括有害内容检测、年龄适配性分类、响应一致性分析、安全对齐微调，以及儿童发展语言建模。除可用于模型训练外，该语料库还可作为可复现的基准测试集，用于评估儿童与AI交互场景中的歧义消解、风险类别标注以及多轮对话安全性。通过为低龄用户标准化安全交互场景，ChildSafeLLM推动了可信AI设计领域的研究，并助力从业者评估生成式模型如何适配儿童的认知与情感需求。（2025-06-10）