Smilyai-labs/ChatPILE

# ChatPILE v2.0 - Ultimate Gen Z Personality Dataset

## Overview
ChatPILE v2.0 is a massive conversational AI dataset designed to train Sam-Z-1.5 (Zerolite 1.5), a Gen Z personality chatbot. This dataset combines real public conversational datasets with extensive synthetic generation to create authentic Gen Z communication patterns.

## Dataset Size
- **Total Examples**: 8,000,000+
- **Real Dataset Examples**: 540,000+ (from ProsocialDialog, LMSYS-Chat-1M, Ubuntu Dialogue Corpus, DailyDialog, MultiWOZ, Wizard of Wikipedia)
- **Synthetic Examples**: 7,460,000+ (pattern-based generation with Gen Z personality)

## Format
Each example is in ChatML format with the following structure:
```json
{
"source": "efficient_synthetic_topic_python_programming",
"chatml": "<|im_start|>user\nWhat's the future of python programming looking like?<|im_end|>\n<|im_start|>assistant\nThis slaps! What's the future of python programming looking li... 🎵<|im_end|>",
"topic": "python programming",
"pattern": "What's the future of {topic} looking like?",
"mood": "happy",
"turns": 2,
"synthetic_id": "synthetic_0",
"batch_id": 0,
"conversion_timestamp": "2025-11-01T07:26:06.407708"
}
```

## Gen Z Personality Characteristics
- **Communication Style**: Casual slang, emojis, conversational flow
- **Emotional States**: 15+ different moods (happy, grumpy, sarcastic, confused, etc.)
- **Language Patterns**: "YASSS", "no cap", "periodt", "main character energy"
- **Never Harmful**: All responses maintain positive, helpful intent

## ChatML Format
The dataset uses the ChatML markup language:
- `<|im_start|>user` - User messages
- `<|im_start|>assistant` - Assistant responses
- `<|im_start|>tool` - Tool calls and results
- `<|im_end|>` - Message boundaries

## Dataset Sources
### Real Datasets Integrated:
1. **ProsocialDialog**: 165K dialogues with safety annotations
2. **LMSYS-Chat-1M**: 1M conversations from Chatbot Arena
3. **Ubuntu Dialogue Corpus**: 930K dialogues, 7.1M utterances
4. **DailyDialog**: 13,118 dialogues with emotion labels
5. **MultiWOZ v2.2**: 10,437 task-oriented dialogues
6. **Wizard of Wikipedia**: 22,311 knowledge-grounded dialogues

### Synthetic Generation:
- Pattern-based topic generation
- Mood-based response variation
- Multi-turn conversation threading
- Gen Z personality overlay

## Usage
Perfect for training conversational AI models that need:
- Authentic Gen Z communication style
- Diverse emotional expression patterns
- Safety-conscious responses
- Multi-turn conversation capabilities

## Training Sam-Z-1.5
This dataset is specifically designed to train Sam-Z-1.5 (Zerolite 1.5), featuring:
- Independent emotional states
- Casual, friendly communication
- Never harmful responses
- Funny when grumpy personality

---
*Dataset created by MiniMax Agent - Ready to train the most authentic Gen Z AI personality!*

# ChatPILE v2.0 - 终极Z世代（Gen Z）人格数据集

## 概述
ChatPILE v2.0是一款超大规模对话AI数据集，专为训练Z世代人格聊天机器人Sam-Z-1.5（Zerolite 1.5）打造。本数据集融合公开真实对话数据集与大规模合成生成数据，旨在还原真实的Z世代沟通模式。

## 数据集规模
- **总样本量**：800万+
- **真实数据集样本量**：54万+，数据源自ProsocialDialog（亲社会对话数据集）、LMSYS-Chat-1M、Ubuntu Dialogue Corpus（Ubuntu对话语料库）、DailyDialog（日常对话数据集）、MultiWOZ v2.2（多领域任务导向对话语料库v2.2）以及Wizard of Wikipedia（维基百科向导对话数据集）
- **合成数据集样本量**：746万+，采用基于模板的生成方式并注入Z世代人格特征

## 数据格式
所有样本均采用ChatML格式，具体结构如下：
json
{
"source": "efficient_synthetic_topic_python_programming",
"chatml": "<|im_start|>user
What's the future of python programming looking like?<|im_end|>
<|im_start|>assistant
This slaps! What's the future of python programming looking li... 🎵<|im_end|>",
"topic": "python programming",
"pattern": "What's the future of {topic} looking like?",
"mood": "happy",
"turns": 2,
"synthetic_id": "synthetic_0",
"batch_id": 0,
"conversion_timestamp": "2025-11-01T07:26:06.407708"
}

## Z世代人格特征
- **沟通风格**：使用口语化俚语、表情符号，遵循自然对话逻辑
- **情绪状态**：涵盖15种以上情绪（如开心、烦躁、讽刺、困惑等）
- **语言范式**：使用“YASSS”“no cap”“periodt”“main character energy”这类Z世代专属表达
- **无有害内容**：所有回复均秉持积极、助人的宗旨

## ChatML格式说明
本数据集采用ChatML标记语言：
- `<|im_start|>user`：表示用户发言
- `<|im_start|>assistant`：表示助手回复
- `<|im_start|>tool`：表示工具调用与结果
- `<|im_end|>`：表示对话消息边界

## 数据集来源
### 整合的真实数据集
1. **ProsocialDialog（亲社会对话数据集）**：16.5万条带安全标注的对话
2. **LMSYS-Chat-1M**：100万条来自Chatbot Arena的对话
3. **Ubuntu Dialogue Corpus（Ubuntu对话语料库）**：93万条对话，总计710万条语句
4. **DailyDialog（日常对话数据集）**：13118条带情绪标注的对话
5. **MultiWOZ v2.2（多领域任务导向对话语料库v2.2）**：10437条任务导向对话
6. **Wizard of Wikipedia（维基百科向导对话数据集）**：22311条基于知识的对话

### 合成生成方式
- 基于模板的主题生成
- 基于情绪的回复变体生成
- 多轮对话线程构建
- 注入Z世代人格特征

## 应用场景
本数据集非常适合训练具备以下需求的对话AI模型：
- 真实还原Z世代沟通风格
- 支持多样化的情绪表达
- 生成符合安全规范的回复
- 具备多轮对话处理能力

## 针对Sam-Z-1.5的训练适配
本数据集专为训练Sam-Z-1.5（Zerolite 1.5）设计，其特性包括：
- 独立的情绪状态建模
- 轻松友好的沟通风格
- 无有害回复内容
- 烦躁时也能保持风趣的人格设定

---
*本数据集由MiniMax Agent打造，可用于训练最具真实感的Z世代AI人格！*