ecommerce-intent
收藏资源简介:
ZeroProof E-commerce Intent 是一个用于支付意图分类的英文文本数据集,专门为训练小型模型以在AI智能体执行操作前验证用户实际请求而构建。数据集包含客户与客服之间的对话,每条对话标注有一个结构化的意图对象,涵盖七种核心类型:消费(spend)、发送(send)、兑换(exchange)、定期支付(recur)、账单(bill)、撤销(reverse)和无意图(none)。数据通过多智能体市场模拟生成:语言模型分别扮演客户和客服角色进行回合制对话,每段对话独立采样人物角色、情境、语气、设备和使用行为,并包含对抗性角色。生成过程采用标签盲法(客户模型仅被告知在购物,而非生成训练样本),标签在后续步骤中根据锁定策略单独标注,且每个数据分割均通过结构性数据门控,确保训练集与测试集零泄漏。数据格式为JSONL,每条记录包含消息序列(含角色和内容)和目标对象(包括意图检测标志、核心类型、详细信息字段、置信度、理由和源消息序列索引)。该数据集设计为领域专用,适用于电子商务和智能体商务场景中的支付意图模型训练与评估。
ZeroProof E-commerce Intent is an English text dataset for payment intent classification, specifically constructed for training small-scale models to verify users' actual requests prior to AI Agent operations. This dataset contains conversations between customers and customer service agents, with each conversation annotated with a structured intent object covering seven core categories: spend, send, exchange, recur, bill, reverse, and none. The dataset is generated via multi-agent market simulation: language models assume the roles of customers and customer service agents respectively for turn-based conversations. Each conversation independently samples character roles, scenarios, conversational tones, devices, and usage behaviors, and incorporates adversarial roles. The generation process employs label blinding: the customer model is only notified that it is participating in a shopping scenario, rather than being informed that it is generating training samples. Labels are separately annotated in subsequent steps using a locking strategy, and each data split undergoes structured data gating to ensure zero data leakage between the training and test sets. The dataset follows the JSONL format, where each record includes a message sequence (comprising speaker roles and corresponding content) and a target object. The target object contains fields such as intent detection flag, core category, detailed information, confidence score, rationale, and source message sequence index. This is a domain-specific dataset tailored for e-commerce and agent-based commerce scenarios, and is designed for training and evaluating payment intent classification models.
数据集概述
数据集名称:ZeroProof E-commerce Intent
发布者:zero-proof-ai
许可证:CC-BY-4.0
语言:英文
数据集大小:10K 到 100K 条样本
任务类别:文本分类
标签:意图分类、支付、电子商务、智能商务、合成数据
数据内容
数据集包含客户与客服之间的对话,每条对话标注了一个结构化意图对象,意图分为七类:
- spend(消费)
- send(发送)
- exchange(兑换/交换)
- recur(循环/定期)
- bill(账单)
- reverse(撤销/退款)
- none(无意图)
每条样本包含:
- messages:对话历史,包含用户和助手的消息序列
- target:意图标注对象,包括:
intent_detected:是否检测到意图core_type:核心意图类型details:该意图的补充字段(如 action、reason_code)confidence:置信度reason:标注理由source_message_seqs:支撑标注的用户消息序号
数据生成方式
数据通过多智能体市场模拟生成,非爬取或模板化。语言模型扮演顾客和客服进行多轮对话,独立采样角色、情境、语气、设备和行为,包括对抗性角色。生成过程对标签不可知(顾客模型仅被告知在购物,不被告知正在生成训练样本),标签在独立阶段依据固定策略分配,且训练集和测试集之间无泄漏。
预期用途
用于训练和评估电子商务及智能商务场景下的支付意图模型。设计上为窄领域、特定任务,非通用指令集,仅支持英文。
基于该数据训练的模型





