README.md

---
license: apache-2.0
dataset_info:
  features:
  - name: version
    dtype: int64
  - name: id
    dtype: string
  - name: mapWidth
    dtype: int64
  - name: mapHeight
    dtype: int64
  - name: usernames
    list: string
  - name: stars
    list: int64
  - name: cities
    list: int64
  - name: cityArmies
    list: int64
  - name: generals
    list: int64
  - name: mountains
    list: int64
  - name: moves
    list:
      list: int64
  splits:
  - name: train
    num_bytes: 424747643
    num_examples: 18803
  download_size: 61449177
  dataset_size: 424747643
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
tags:
- games
- replays
- generalsio
---

# ⚔️ Generals.io High-Rank Replay Dataset 🌟

## Overview

This dataset contains a curated collection of 1v1 game replays from the online strategy game [generals.io](httphttp://generals.io), specifically designed for training high-level reinforcement learning agents 🤖.

* 🏆 **High-Quality Matches:** Includes games where at least one participant had a star rating of **70 or higher** 📈, ensuring a baseline of quality and strategic depth.
* ✅ **Clean Data:** Carefully filtered to remove outliers, games with AFK players, and trolls, providing a clean and focused dataset for agent training.
* 🚀 **Proven Effectiveness:** An agent trained exclusively on this dataset has been shown to be capable of reaching a **60-star rating**, demonstrating the sufficiency and quality of the provided replays.

### Game Version

All replays were recorded on the game patch featuring an **alternating priority** turn system. However, extensive testing has shown that agents trained on this data generalize effectively to the newer priority system without requiring further training.

## Dataset Structure & Documentation

The dataset is composed of individual JSON files, where each file represents a single game replay. For a detailed technical description of the replay format, please refer to the official (archived) documentation:

➡️ **[http://dev.generals.io/replays (via Archive.org)](https://web.archive.org/web/20190917202205/http://dev.generals.io/replays)**

The structure of each JSON object is as follows:

| Field         | Type                      | Description                                                                                             |
|---------------|---------------------------|---------------------------------------------------------------------------------------------------------|
| `version`     | `Integer`                 | The version number of the replay format.                                                                |
| `id`          | `String`                  | A unique identifier for the replay.                                                                     |
| `mapWidth`    | `Integer`                 | The width of the game map in tiles.                                                                     |
| `mapHeight`   | `Integer`                 | The height of the game map in tiles.                                                                    |
| `usernames`   | `List[String]`            | A list of the usernames for the players in the game. The index corresponds to the player's turn order.  |
| `stars`       | `List[Integer]`           | The star rating of each player at the time the game was played.                                         |
| `cities`      | `List[Integer]`           | A list of tile indices representing the locations of all neutral cities at the start of the game.       |
| `cityArmies`  | `List[Integer]`           | The initial army count for each corresponding city in the `cities` list.                                |
| `generals`    | `List[Integer]`           | The starting tile indices for each player's general (home base).                                        |
| `mountains`   | `List[Integer]`           | A list of tile indices for all impassable mountain tiles on the map.                                    |
| `moves`       | `List[List[Integer]]`     | The core sequence of game actions. Each inner list is a move tuple, typically representing `[player_index, start_tile, end_tile, is_50%_move, turn_number]`. |
| `afks`        | `List[]`                  | A list that would contain information on AFK players. This has been filtered and will always be empty.    |

## Usage Example

This dataset is hosted on the Hugging Face Hub and can be loaded directly using the `datasets` library. Here is a showcase of how to load the dataset and inspect its contents.

```python
from datasets import load_dataset

# 1. Load the dataset directly from the Hugging Face Hub
#    You may need to log in first: huggingface-cli login
print("Loading dataset...")
dataset = load_dataset("strakammm/generals_io_replays")

# The dataset is loaded into a DatasetDict, we'll use the 'train' split
train_dataset = dataset['train']

# 2. Print the total number of replays in the dataset
print(f"\nTotal number of replays: {len(train_dataset)}")

# 3. Iterate over the first few replays to showcase the data
print("\n--- Showcasing first 5 replays ---")
for i in range(5):
    replay = train_dataset[i]
    
    # Get the players and the number of moves
    players = replay['usernames']
    num_moves = len(replay['moves'])
    
    print(f"\nReplay {i+1}:")
    print(f"  Players: {players[0]} vs {players[1]}")
    print(f"  Total moves: {num_moves}")

print("\n------------------------------------")
```


## Citation

If you use this dataset in your research, please cite the following paper:

```bibtex
@misc{generals_rl,
      author    = {Matej Straka, Martin Schmid},
      title     = {Artificial Generals Intelligence: Mastering Generals.io with Reinforcement Learning},
      year      = {2025},
      eprint    = {2507.06825},
      archivePrefix = {arXiv},
      primaryClass = {cs.LG},
}