Datasets:

webxos
/

arachnid_RL

language:
  - en
task_categories:
  - reinforcement-learning
  - tabular-classification
  - tabular-to-text
tags:
  - human-demonstrations
  - atari-like
  - browser-game
  - rl-dataset
  - games
  - gym
  - reinforcement-learning
  - robotics
  - atari
  - 2d
  - RLHF
  - RL
  - video-game
license: mit
pretty_name: ARACHNID RL Dataset
size_categories:
  - n<1K
config_name: default

      ___           ___           ___           ___           ___           ___                               
     /\  \         /\  \         /\  \         /\__\         /\  \         /\  \                     _____    
    /::\  \       /::\  \       /::\  \       /:/  /         \:\  \        \:\  \       ___         /::\  \   
   /:/\:\  \     /:/\:\__\     /:/\:\  \     /:/  /           \:\  \        \:\  \     /\__\       /:/\:\  \  
  /:/ /::\  \   /:/ /:/  /    /:/ /::\  \   /:/  /  ___   ___ /::\  \   _____\:\  \   /:/__/      /:/  \:\__\ 
 /:/_/:/\:\__\ /:/_/:/__/___ /:/_/:/\:\__\ /:/__/  /\__\ /\  /:/\:\__\ /::::::::\__\ /::\  \     /:/__/ \:|__|
 \:\/:/  \/__/ \:\/:::::/  / \:\/:/  \/__/ \:\  \ /:/  / \:\/:/  \/__/ \:\~~\~~\/__/ \/\:\  \__  \:\  \ /:/  /
  \::/__/       \::/~~/~~~~   \::/__/       \:\  /:/  /   \::/__/       \:\  \        ~~\:\/\__\  \:\  /:/  / 
   \:\  \        \:\~~\        \:\  \        \:\/:/  /     \:\  \        \:\  \          \::/  /   \:\/:/  /  
    \:\__\        \:\__\        \:\__\        \::/  /       \:\__\        \:\__\         /:/  /     \::/  /   
     \/__/         \/__/         \/__/         \/__/         \/__/         \/__/         \/__/       \/__/

ARACHNID RL Dataset

This dataset contains reinforcement learning transitions collected from human gameplay of ARACHNID RL, a 2D Atari-inspired space shooter. It contains about 2,831 samples of human gameplay data from a simple Atari-inspired space shooter game.

Players control a spider-like ship to shoot asteroids and aliens while collecting diamonds. To build your own datasets download the ARACHNID RL file in the /gym/ folder of this repo. Play the game and build your own datasets based off of input data. The game features desktop keyboard and mobile oneclick browser support.

The dataset is designed for RL research, such as training agents via imitation learning or behavioral cloning from human demonstrations.

Game Description

Players control a spider-like ship to destroy asteroids and aliens while collecting diamonds. Unlike large-scale RL datasets focused on real-world robotics or massive web-derived preferences, arachnid_RL emphasizes simplicity and interpretability. Its modest scale (1K–10K samples) makes it suitable for rapid prototyping, educational purposes, or benchmarking algorithms on human-like decision-making in a dynamic but low-dimensional environment.

Dataset Structure

The main dataset is in data/train.jsonl in JSON Lines format. 1.83 MB, stored primarily as a JSON Lines file (train.jsonl), with an auto-converted Parquet version for efficient loading.

Each entry represents a single transition, including timestamp, session/player ID, event type (e.g., shoot, move, game_start, destroy_alien), action taken (e.g., left, right, shoot), reward (e.g., +15 for collecting diamonds), done flag, current state (as JSON with position, velocity, score, lives, nearby objects, etc.), next state, and event details.

Data Format

Each line in data/train.jsonl is a JSON object with:

state: Game state (JSON string containing position, velocity, lives, score, nearby objects)
action: Player action (left, right, up, down, shoot, boost, none)
reward: Immediate reward
next_state: Next game state (JSON string)
done: Episode termination flag
event_type: Type of event
event_details: Additional metadata (JSON string)

Citation

@misc{arachnid_rl,
  title = {ARACHNID RL Dataset},
  author = {WebXOS},
  year = {2026}
}

License

MIT License