Datasets:

elizaos
/

babylon-game-data

license: mit
task_categories:
  - reinforcement-learning
  - game-simulation
  - agent-training
tags:
  - babylon
  - prediction-markets
  - game-worlds
  - agent-trajectories
  - offline-simulation
size_categories:
  - 10K<n<100K

elizaos/babylon-game-data

Dataset Description

Complete Babylon game data for reinforcement learning and offline simulation.

Version: 1.0.0
Collected: 2025-11-16T04:18:42.175Z
Game Worlds: 2
Agent Trajectories: 20
Benchmarks: 4

What's Included

1. Complete Game Worlds

Prediction market scenarios
30-day timelines with events
NPC conversations and interactions
Feed posts and social dynamics
Ground truth outcomes

2. Agent Trajectories

Complete agent decision sequences
LLM calls (prompts and responses)
Game environment at each step
Actions taken and outcomes
Rewards and ground truth

3. Benchmark Results

Model performance evaluations
Comparison to baselines
Detailed metrics

Data Organization

By Month

by-month/
  2025-10.json  - October 2025 data
  2025-11.json  - November 2025 data
  2025-12.json  - December 2025 data
  ...

Each month file contains:

Game worlds generated that month
Agent trajectories from that month
Benchmark results from that month

Offline Simulation

This dataset enables offline, faster-than-real-time simulation:

# Download dataset
from datasets import load_dataset
dataset = load_dataset("elizaos/babylon-game-data")

# Load into Babylon offline simulator
bun run scripts/run-offline-simulation.ts \
  --data=path/to/downloaded/data.json \
  --fast-forward \
  --agent=my-agent

Use Cases

RL Training - Train agents on historical gameplay
Model Evaluation - Test agents on past scenarios
Offline Development - Develop without live system
Research - Analyze agent behavior and game dynamics
Faster Testing - Run simulations at high speed

Data Format

Game World

{
  "worldId": "...",
  "month": "2025-11",
  "question": "Will Bitcoin reach $100k?",
  "outcome": true,
  "timeline": [ /* 30 days of events */ ],
  "npcs": [ /* NPC data */ ],
  "events": [ /* All events */ ],
  "feedPosts": [ /* Social feed */ ]
}

Agent Trajectory

{
  "trajectoryId": "...",
  "month": "2025-11",
  "steps": [
    {
      "environment_state": { /* game state */ },
      "llm_calls": [ /* agent decisions */ ],
      "action": { /* what agent did */ },
      "reward": 50
    }
  ],
  "totalReward": 1500,
  "finalPnL": 1500
}

Citation

@dataset{babylon_game_data_2025,
  title = {Babylon Game Data - Complete RL Dataset},
  author = {Babylon Labs},
  year = {2025},
  url = {https://huggingface.co/datasets/elizaos/babylon-game-data}
}

License

MIT