nortem/marl-gpt-datasets
收藏Hugging Face2026-03-02 更新2026-03-29 收录
下载链接:
https://hf-mirror.com/datasets/nortem/marl-gpt-datasets
下载链接
链接失效反馈官方服务:
资源简介:
---
license: mit
pretty_name: MARL-GPT Datasets
tags:
- reinforcement-learning
- multi-agent
- offline-rl
- trajectories
task_categories:
- reinforcement-learning
---
# MARL-GPT Datasets
Offline expert trajectories from **“MARL-GPT: Foundation Model for Multi-Agent Reinforcement Learning”**.
## Environments
This dataset includes trajectories from the three evaluation domains used in MARL-GPT: **SMACv2** (StarCraft multi-agent combat), **Google Research Football (GRF)**, and **POGEMA** (partially observable multi-agent pathfinding on grids).
## Format
Trajectories are stored sequentially (no shuffling). Use the `done` flag to split the stream into per-agent trajectory segments.
Each transition is:
- `(obs, act, rew, done)`
Training-ready data is provided as a flattened dict of tensors:
- `obs`: `(B, n_obs_feat)`
- `act`: `(B,)`
- `rew`: `(B,)`
- `done`: `(B,)` (optional)
- `info_battle_won`: `(B,)` (optional)
- `action_mask`: `(B, n_act)` (optional)
`B` is the number of transitions in the flattened stream.
license: MIT许可证(mit)
pretty_name: MARL-GPT 数据集
tags:
- 强化学习(reinforcement learning)
- 多智能体(multi-agent)
- 离线强化学习(offline-rl)
- 轨迹(trajectories)
task_categories:
- 强化学习(reinforcement learning)
# MARL-GPT 数据集
本数据集源自论文**《MARL-GPT:面向多智能体强化学习的基础模型》**中的离线专家轨迹数据。
## 数据集涵盖的环境
本数据集包含MARL-GPT论文中使用的三类评估环境对应的轨迹数据:**SMACv2(星际争霸多智能体对战,StarCraft multi-agent combat)**、**谷歌研究足球(Google Research Football, GRF)** 以及 **POGEMA(网格环境下的部分可观测多智能体寻路,partially observable multi-agent pathfinding on grids)**。
## 数据格式
轨迹数据按序列存储(未进行打乱操作),可通过`done`标记将数据流切分为单智能体轨迹片段。
每条过渡样本的格式为:
- `(obs, act, rew, done)`
可直接用于训练的数据以扁平化张量字典形式提供:
- `obs`:形状为`(B, n_obs_feat)`的张量
- `act`:形状为`(B,)`的张量
- `rew`:形状为`(B,)`的张量
- `done`:形状为`(B,)`的张量(可选字段)
- `info_battle_won`:形状为`(B,)`的张量(可选字段,战斗胜利标记)
- `action_mask`:形状为`(B, n_act)`的张量(可选字段,动作掩码)
其中`B`为扁平化数据流中过渡样本的总数量。
提供机构:
nortem


