rescommons/Full-Ecom-Chatbot-Dataset
收藏Hugging Face2026-03-27 更新2026-03-29 收录
下载链接:
https://hf-mirror.com/datasets/rescommons/Full-Ecom-Chatbot-Dataset
下载链接
链接失效反馈官方服务:
资源简介:
---
annotations_creators:
- machine-generated
- expert-generated
language_creators:
- machine-generated
- found
language:
- en
license: mit
multilinguality:
- monolingual
size_categories:
- 10K<n<100K
source_datasets:
- original
task_categories:
- question-answering
- text-generation
- other
task_ids:
- dialogue-modeling
- open-domain-qa
pretty_name: E-commerce Chatbot Training Data
tags:
- ecommerce
- chatbot
- tool-use
- customer-support
- retail
- conversational-ai
---
# E-commerce Chatbot Training Data
A curated, multi-source dataset for training and evaluating e-commerce conversational AI systems. It covers a broad range of customer intents — from product discovery and order management to returns, tool-augmented responses, and RAG-grounded Q&A — across 16+ product domains.
## Dataset Summary
| Split | Records |
|-------|---------|
| Train | 35,213 |
| Test | 8,818 |
| **Total** | **44,031** |
The train/test split uses **prompt-group-level stratified sampling** on `source × response_type × intent × difficulty` to guarantee identical distributions across both splits with zero prompt contamination between train and test.
---
## Sources
| Source | Records | Response Types | Domains | Intents |
|--------|---------|----------------|---------|---------|
| `synthetic_api_generated` | 3,933 | text, tool_call, mixed | 12 | 19 |
| `asos_ecom_dataset` | 2,000 | text | fashion | similarity_search |
| `bitext_customer_support` | 5,000 | tool_call, mixed | general | 6 |
| `bitext_retail_ecom` | 4,998 | text, tool_call | general | multiple |
| `amazon_reviews_2023_*` | 23,100 | text | 16 | 4 |
| `amazon_meta_2023_*` | 5,000 | text | 9 | 4 |
---
## Schema
| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique record ID (e.g. `ecomm_a1b2c3`) |
| `source` | string | Origin dataset/pipeline |
| `group` | string | Response group: `A` (tool_call), `B` (text), `C` (mixed) |
| `difficulty` | int | Task difficulty: `1` (easy) to `3` (hard) |
| `system` | string | System prompt given to the assistant |
| `history` | string (JSON) | Prior conversation turns `[{"role": ..., "content": ...}]` |
| `prompt` | string | Current user message |
| `context` | string (JSON) | Retrieved docs, user profile, cart/order state |
| `tools` | string (JSON) | Available tool/function definitions |
| `response_type` | string | `text`, `tool_call`, or `mixed` |
| `response` | string | Ground-truth assistant response |
| `language` | string | ISO language code (e.g. `en`) |
| `locale` | string | Locale (e.g. `en-US`) |
| `annotator` | string | Annotation source (e.g. `gemini_synthetic`, `bitext`, `amazon_user`) |
| `quality_score` | float | Annotation quality score (0–1) |
| `domain` | string | Product domain (e.g. `electronics`, `fashion`, `grocery_food`) |
| `intent_category` | string | High-level intent category (e.g. `product_discovery`, `order_management`) |
| `intent` | string | Fine-grained intent (19 values, e.g. `order_status`, `return_refund`) |
| `sub_intent` | string | Further sub-intent (e.g. `track_delivery`, `refund_timeline`) |
| `capability` | string | Model capability tag (where applicable) |
| `test_tier` | string | Evaluation tier tag (where applicable) |
---
## Intents
The dataset covers 19 intents across 7 high-level categories:
| Category | Intents |
|----------|---------|
| Product Discovery | `product_search`, `product_detail_qa`, `product_comparison`, `similarity_search`, `bundle_suggestions`, `gift_recommendation`, `personalized_recommendations` |
| Order Management | `order_status`, `order_cancellation`, `reorder_assistance` |
| Returns & Exchanges | `return_refund`, `exchange_request` |
| Cart & Checkout | `cart_management`, `payment_issues` |
| Customer Support | `complaint_handling`, `human_handoff`, `faq_answering` |
| Account | `account_management` |
| Inventory | `stock_availability` |
---
## Product Domains
`appliances`, `beauty`, `books_media`, `electronics`, `fashion`, `gaming`, `garden_outdoor`, `grocery_food`, `home_kitchen`, `industrial`, `pet_supplies`, `sports_outdoors`, `automotive`, `baby`, `health`, `office`, `toys_games`
---
## Usage
```python
from datasets import load_dataset
ds = load_dataset("V1rtucious/ecom-chatbot-train-data")
train = ds["train"]
test = ds["test"]
# Filter by response type
tool_call_examples = train.filter(lambda x: x["response_type"] == "tool_call")
# Filter by intent
order_queries = train.filter(lambda x: x["intent"] == "order_status")
```
---
## Split Methodology
Both splits were produced using **prompt-group-level stratified sampling** to ensure zero contamination, maximum variance, and minimum bias:
- **Stratification key:** `source | response_type | intent | difficulty`
- **Splitting unit:** unique `(source, prompt)` groups — all records sharing a prompt are assigned atomically to one split
- **40,949 prompt groups** across 44,031 records; 3,082 records share a prompt with at least one other record
- **Fallback cascade** for rare strata (< 5 groups): drops `difficulty`, then drops to `source` only
- **113 unique strata** | **Random seed:** 42 (reproducible)
- **Prompt contamination between splits: 0** (verified post-split)
Distribution drift between train and test is < 0.35% across all key columns.
---
## License
This dataset is released under the **MIT License**. Individual source data may carry additional terms from their original providers (Amazon, ASOS, Bitext).
annotations_creators:
- 机器生成(machine-generated)
- 专家生成(expert-generated)
language_creators:
- 机器生成(machine-generated)
- 公开获取(found)
language:
- 英语(en)
license: MIT许可证(MIT)
multilinguality:
- 单语言(monolingual)
size_categories:
- 10K<n<100K
source_datasets:
- 原始数据集(original)
task_categories:
- 问答(question-answering)
- 文本生成(text-generation)
- 其他(other)
task_ids:
- 对话建模(dialogue-modeling)
- 开放域问答(open-domain-qa)
pretty_name: 电子商务聊天机器人训练数据(E-commerce Chatbot Training Data)
tags:
- 电子商务(ecommerce)
- 聊天机器人(chatbot)
- 工具使用(tool-use)
- 客户支持(customer-support)
- 零售(retail)
- 对话式AI(conversational-ai)
# 电子商务聊天机器人训练数据集(E-commerce Chatbot Training Data)
本数据集为经精心整理的多源数据集,用于训练与评估电子商务对话式AI(conversational AI)系统。其覆盖16余个产品领域,涵盖丰富的用户意图,从商品发现、订单管理,到退换货、工具增强型回复,以及基于检索增强生成(Retrieval-Augmented Generation,RAG)的问答任务。
## 数据集概览
| 拆分布局 | 样本数量 |
|-------|---------|
| 训练集 | 35,213 |
| 测试集 | 8,818 |
| **总计** | **44,031** |
本次训练集与测试集的划分采用**基于提示词组的分层抽样(prompt-group-level stratified sampling)**,抽样维度为`数据源 × 回复类型 × 意图 × 难度`,以确保两个拆分集合的分布完全一致,且训练集与测试集之间无任何提示词污染(prompt contamination)。
---
## 数据来源
| 数据源 | 样本数量 | 回复类型 | 覆盖领域 | 意图数量 |
|--------|---------|----------------|---------|---------|
| `synthetic_api_generated` | 3,933 | 文本、工具调用、混合 | 12个 | 19种 |
| `asos_ecom_dataset` | 2,000 | 仅文本 | 时尚领域 | 相似度搜索 |
| `bitext_customer_support` | 5,000 | 工具调用、混合 | 通用领域 | 6种 |
| `bitext_retail_ecom` | 4,998 | 文本、工具调用 | 通用领域 | 多种 |
| `amazon_reviews_2023_*` | 23,100 | 仅文本 | 16个领域 | 4种 |
| `amazon_meta_2023_*` | 5,000 | 仅文本 | 9个领域 | 4种 |
---
## 数据 Schema
| 字段名 | 数据类型 | 字段说明 |
|-------|------|-------------|
| `id` | 字符串(string) | 唯一记录ID(示例:`ecomm_a1b2c3`) |
| `source` | 字符串 | 原始数据集/处理流水线来源 |
| `group` | 字符串 | 回复分组:`A`(工具调用)、`B`(文本回复)、`C`(混合回复) |
| `difficulty` | 整数(int) | 任务难度:`1`(简单)至`3`(困难) |
| `system` | 字符串 | 分配给助手的系统提示词 |
| `history` | JSON格式字符串 | 历史对话轮次,格式为`[{"role": ..., "content": ...}]` |
| `prompt` | 字符串 | 当前用户输入的消息 |
| `context` | JSON格式字符串 | 检索到的文档、用户个人资料、购物车/订单状态 |
| `tools` | JSON格式字符串 | 可用工具/函数定义 |
| `response_type` | 字符串 | 回复类型:`text`(文本)、`tool_call`(工具调用)或`mixed`(混合) |
| `response` | 字符串 | 助手的真实标注回复 |
| `language` | 字符串 | ISO语言代码(示例:`en`) |
| `locale` | 字符串 | 区域设置(示例:`en-US`) |
| `annotator` | 字符串 | 注释来源(示例:`gemini_synthetic`、`bitext`、`amazon_user`) |
| `quality_score` | 浮点数(float) | 注释质量评分(范围0至1) |
| `domain` | 字符串 | 产品领域(示例:`electronics`(电子产品)、`fashion`(时尚)、`grocery_food`(食品杂货)) |
| `intent_category` | 字符串 | 高级意图类别(示例:`product_discovery`(商品发现)、`order_management`(订单管理)) |
| `intent` | 字符串 | 细粒度意图(共19种,示例:`order_status`(订单状态查询)、`return_refund`(退款退货)) |
| `sub_intent` | 字符串 | 进一步细分的子意图(示例:`track_delivery`(物流追踪)、`refund_timeline`(退款进度查询)) |
| `capability` | 字符串 | 模型能力标签(适用于相关场景) |
| `test_tier` | 字符串 | 评估层级标签(适用于相关场景) |
---
## 意图分类
本数据集涵盖7个高级类别下的19种细粒度意图:
| 高级意图类别 | 细粒度意图 |
|----------|---------|
| 商品发现(Product Discovery) | `product_search`(商品搜索)、`product_detail_qa`(商品详情问答)、`product_comparison`(商品对比)、`similarity_search`(相似度搜索)、`bundle_suggestions`(套餐推荐)、`gift_recommendation`(礼品推荐)、`personalized_recommendations`(个性化推荐) |
| 订单管理(Order Management) | `order_status`(订单状态查询)、`order_cancellation`(订单取消)、`reorder_assistance`(重新下单协助) |
| 退换货与换货(Returns & Exchanges) | `return_refund`(退款退货)、`exchange_request`(换货申请) |
| 购物车与结账(Cart & Checkout) | `cart_management`(购物车管理)、`payment_issues`(支付问题) |
| 客户支持(Customer Support) | `complaint_handling`(投诉处理)、`human_handoff`(转人工客服)、`faq_answering`(常见问题解答) |
| 账户管理(Account) | `account_management`(账户管理) |
| 库存查询(Inventory) | `stock_availability`(库存可用性查询) |
---
## 产品覆盖领域
`appliances`(家电)、`beauty`(美妆)、`books_media`(图书音像)、`electronics`(电子产品)、`fashion`(时尚服饰)、`gaming`(游戏)、`garden_outdoor`(园艺户外)、`grocery_food`(食品杂货)、`home_kitchen`(家居厨房)、`industrial`(工业用品)、`pet_supplies`(宠物用品)、`sports_outdoors`(运动户外)、`automotive`(汽车用品)、`baby`(母婴)、`health`(健康保健)、`office`(办公用品)、`toys_games`(玩具游戏)
---
## 使用方法
python
from datasets import load_dataset
ds = load_dataset("V1rtucious/ecom-chatbot-train-data")
train = ds["train"]
test = ds["test"]
# 按回复类型筛选
tool_call_examples = train.filter(lambda x: x["response_type"] == "tool_call")
# 按意图筛选
order_queries = train.filter(lambda x: x["intent"] == "order_status")
---
## 拆分方法说明
本次拆分采用**基于提示词组的分层抽样**方法,以确保无提示词污染、样本方差最大化与偏差最小化:
- **分层维度**:`数据源 | 回复类型 | 意图 × 难度`
- **拆分单元**:唯一的`(数据源, 提示词)`组 — 所有共享同一提示词的记录会被整体分配至同一个拆分集合
- 44,031条样本共包含40,949个提示词组;其中3,082条样本与至少一条其他样本共享同一提示词
- **稀有分层降级策略**:对于样本量小于5个组的稀有分层,依次移除`难度`维度,最终仅保留`数据源`维度
- 共113个唯一分层 | **随机种子**:42(可复现)
- 拆分集合间的提示词污染率为0(拆分后已验证)
训练集与测试集在所有关键列上的分布漂移率均小于0.35%。
---
## 许可证
本数据集采用**MIT许可证**进行发布。各原始数据源可能附带其原始提供方(Amazon(亚马逊)、ASOS、Bitext)的额外条款。
提供机构:
rescommons



