---
title: Structured-GameNPCBehavior-InteractionDataset
canonical_url: "https://www.modelscope.cn/datasets/LoongSerpent9Realms/Structured-GameNPCBehavior-InteractionDataset"
md_url: "https://www.modelscope.cn/datasets/LoongSerpent9Realms/Structured-GameNPCBehavior-InteractionDataset.md"
repository: LoongSerpent9Realms/Structured-GameNPCBehavior-InteractionDataset
chinese_name: "结构化 游戏NPC 行为和交互数据集"
last_updated: 2026-01-19
license: "CC BY 4.0"
storage_size: "14 MB"
downloads: 97
stars: 1
---

# Structured-GameNPCBehavior-InteractionDataset

> Structured-GameNPCBehavior-InteractionDataset - LoongSerpent9Realms 在 ModelScope 开源的数据集。结构化NPC运行时行为决策数据集（NavMesh v2）

LoongSerpent9Realms/Structured-GameNPCBehavior-InteractionDataset 是 ModelScope 魔搭社区上的数据集，存储大小 14 MB，采用 CC BY 4.0 许可。

- **Repository**: LoongSerpent9Realms/Structured-GameNPCBehavior-InteractionDataset
- **License**: CC BY 4.0
- **Storage size**: 14 MB
- **Downloads**: 97
- **Stars**: 1
- **Last updated**: 2026-01-19

Source: https://www.modelscope.cn/datasets/LoongSerpent9Realms/Structured-GameNPCBehavior-InteractionDataset

---

# 结构化NPC运行时行为决策数据集（NavMesh v2）

## 一、数据集简介
本数据集用于训练与评估面向 Unity 3D 游戏环境的 NPC 运行时行为规划模型（Runtime NPC Planner）。  
数据以任务型对话（Task-oriented Dialogue）的形式构建，强调：

- 世界状态持续同步（tick）
- POI 映射版本控制（poi_catalog_version）
- 动作严格白名单约束（AllowedActions）
- 计划锁语义（plan_lock），用于避免重复规划和路径抖动

模型的目标是根据角色背景与当前世界状态，生成合法、可执行、结构化的 JSON 行为计划。

---

## 二、任务类型
- 领域：NLP  
- 任务：Conversational（任务型对话）  
- 子类型：Task-QA / Action Planning  
- 输出形式：严格结构化 JSON（非自然语言）

---

## 三、数据结构

每条样本采用 ChatML 格式，包含三部分：

### 1. System
约束模型为 NPC 规划器，仅允许输出合法 JSON 与白名单动作。

### 2. User
输入结构化世界状态：
- tick（世界版本）
- 地图与 POI 版本
- 玩家位置
- NPC 当前状态（位置、HP、Stamina、当前目标）
- nearby_pois（本轮允许导航目标）
- allowed_actions（动作与参数 schema）

### 3. Assistant
输出：
- plan_id
- goal / intent
- actions（只能使用白名单 + nearby_pois）
- plan_lock（时间锁或目标锁）
- reasoning_brief（调试用）

---

## 四、适用场景
本数据集适用于：

- 游戏 NPC 自主行为规划
- 多智能体任务决策建模
- 运行时 AI 行为生成（Unity / Unreal）
- Agent Planning / Tool-Calling LLM 微调
- 强约束结构化输出任务

---

## 五、数据规模
| 划分 | 数量 |
|------|------|
| Train | 2250 |
| Val | 125 |
| Test | 125 |
| Total | 2500 |

---

## 六、特点
- 强结构化输入输出（无自由文本）
- 明确动作 schema，避免模型幻觉
- 引入计划锁（Plan Lock）模拟持续决策
- POI 局部可见机制（nearby_pois）
- 完全贴合 3D NavMesh 游戏运行时逻辑
