eth-sri/json-mode-eval-extended

这是一个衡量大型语言模型从自然语言中提取数据并遵循JSON Schema能力的数据集。它通过对NousResearch的json-mode-eval数据集进行清洗和标准化，生成了eth-sri的json-mode-eval-cleaned数据集，确保每个模式都执行非空约束并禁止在顶层添加额外的键。然后，对Gemini 2.5 Pro模型进行了额外的样本提示，并过滤出符合给定模式的输出。

This is a dataset that measures LLM capabilities at extracting data from natural language following a JSON Schema. It was generated by cleaning and normalizing the NousResearch json-mode-eval dataset to create the eth-sri json-mode-eval-cleaned dataset, ensuring that every schema enforces non-empty constraints and allows no additional keys at the top level. Additional prompts were provided to the Gemini 2.5 Pro model, and outputs that are valid according to the given schema were filtered.