Mightypeacock/european-flora-fungi-thinking

一个精心策划的数据集，包含377个多轮对话，覆盖39种欧洲植物和蘑菇物种，分为14个易混淆组（常被误认为彼此的物种）。该数据集设计用于微调视觉语言模型（如Gemma 3、LLaVA、Qwen-VL），带有思考轨迹（`<think>...</think>`），以执行以下任务：1) 从照片中进行物种识别，并提供逐步推理；2) 迭代诊断缩小——模型提出候选物种，提出有针对性的问题，用户提供证据，模型更新置信度分数并排除物种，直到达到识别；3) 易混淆物种区分——系统地排除危险的混淆；4) 安全评估——可食用性分类并提供适当警告；5) 错误识别纠正——捕捉并纠正危险的识别错误；6) 人类反馈循环——用户审查置信度评分卡，同意或不同意，模型进行修订。

A curated dataset of 377 multi-turn conversations covering 39 European plant and mushroom species organized into 14 confusion groups (species commonly mistaken for each other). This dataset is designed for fine-tuning vision-language models (Gemma 3, LLaVA, Qwen-VL) with thinking traces (`<think>...</think>`) to perform: 1) Species identification from photographs with step-by-step reasoning; 2) Iterative diagnostic narrowing — model proposes candidates, asks targeted questions, user provides evidence, model updates confidence scores and eliminates species until identification is reached; 3) Lookalike disambiguation — systematically ruling out dangerous confusions; 4) Safety assessment — edibility classification with appropriate warnings; 5) Misidentification correction — catching and correcting dangerous ID mistakes; 6) Human-in-the-loop feedback — user reviews confidence scorecard, agrees/disagrees, model revises.