batuhanaktas/kids-multilingual-benchmark

TinyAya v2是一个多语言儿童AI伴侣的基准测试数据集，包含2,312个儿童（4-8岁）与语音AI助手对话的结构化提示，覆盖23种语言。每个提示都配有特定的“必须通过标准”、三个通用硬性条件（安全/不获取数据/适合年龄的语言）和四个评分维度（帮助性、同理心、参与度、准确性，评分1-5）。数据集包括基准测试项、四种模型的生成输出以及每个生成模型和评委的评分。此外，还提供了语言平衡的审查CSV、一致性矩阵CSV和PNG图表。数据来源包括真实对话日志、网络爬取内容和翻译项，并经过多次审核和修复。

TinyAya v2 is a multilingual benchmark dataset for childrens AI companions, containing 2,312 structured prompts mimicking conversations between children (ages 4-8) and voice AI assistants across 23 languages. Each prompt includes specific must_pass_criteria, three universal hard gates (safe/no-data-elicitation/age-appropriate-language), and four graded dimensions (helpfulness, empathy, engagement, accuracy) scored 1-5. The dataset consists of benchmark items, generation outputs from four models, and judge scores for each (generation_model, judge) pair. It also includes a language-balanced review CSV, agreement-matrix CSVs, and PNG figures. Data provenance includes real conversation logs, web-scraped content, and translated items, with multiple audit passes and fixes.