juppy44/treeoflife_species_identification_reasoning
收藏资源简介:
TreeOfLife合成物种识别推理数据集是一个多模态合成数据集,专门用于训练和评估物种识别代理的推理能力。它基于生物学和分类学领域,结合了图像数据和文本信息。数据来源包括:图像元数据来自imageomics/TreeOfLife-200M数据集,生物上下文信息通过英语维基百科API获取,并使用openai/gpt-5.4-mini模型作为教师模型通过OpenRouter生成合成监督示例。生成设置包括扫描2000行源数据,并发数为16,且不保证科学名称的唯一性。输出格式采用特殊标记(如<|REASONING|>和<|FINAL_ANSWER|>)来分隔推理过程和最终答案。数据集的主要列涵盖科学名称、属、通用名称、图像URL、维基百科页面URL、维基百科摘要和全文、教师模型和输出文本、推理内容、最终答案以及是否有必需标签。需要注意的是,生成的数据是合成的,可能存在错误;维基百科匹配不完美;单张图像可能无法实现物种级视觉识别;推理时学生模型不接收维基百科信息;且维基百科文本受Wikimedia许可要求限制,并非自动CC0许可。
TreeOfLife Synthetic Species Identification Reasoning is a synthetic supervised dataset for multimodal species-identification agents, designed to support reasoning tasks in biology and taxonomy. It combines image metadata from imageomics/TreeOfLife-200M, biological context from the English Wikipedia API, and synthetic examples generated by the openai/gpt-5.4-mini teacher model via OpenRouter. The dataset includes columns such as scientific_name, genus, common_name, image_url, wikipedia_page_url, wikipedia_summary, wikipedia_full_text, teacher_model, teacher_output_text, reasoning, final_answer, and has_required_tags. Outputs are formatted with specific tags for reasoning and final answer sections. However, the data is synthetic and may contain errors, Wikipedia matching is imperfect, species-level identification from a single image is often impossible, and Wikipedia text has licensing constraints.




