milkkarten/pokechamp

Pokémon Champion

This is the implementation for the paper "PokéChamp: an Expert-level Minimax Language Agent for Competitive Pokémon"

Requirements:

conda create -n pokechamp python=3.12
conda activate pokechamp
pip install -r requirements.txt

Configuration

Configuring a Local Pokémon Showdown Server (Battle Engine)

1. Install Node.js v10+.
2. Clone the Pokémon Showdown repository and set it up:

git clone https://github.com/smogon/pokemon-showdown.git
cd pokemon-showdown
# Optional: All repo features were tested with the following showdown version.
# git reset --hard dd4b004e54d4ef8c66c8b583a8fa64b020574727
npm install
cp config/config-example.js config/config.js
node pokemon-showdown start --no-security

Enter "http://localhost:8000/" in your browser.

Reproducing Paper Results

Gen 9 OU Battles

To evaluate a group of agents on Gen 9 OU format:

python evaluate_gen9ou.py

This script will run battles between PokéChamp and the baseline bots, including PokéLLMon, Abyssal Bot, and others. It will output win rates, Elo ratings, and average number of turns per battle.

Additional Experiments

Battle Any Agent Against Any Agent Locally

python local_1v1.py 

Battle Against Any Agent Locally

First, log into your other account manually on the local server, choosing "[Gen 9] Random Battle".

python human_agent_1v1.py 

Battle Against Ladder Players on Pokémon Showdown

Register an account on https://play.pokemonshowdown.com/ and get your password.

Open and log in: https://play.pokemonshowdown.com/

python showdown_ladder.py --USERNAME $USERNAME --PASSWORD $PASSWORD # fill in your username and password for PokéChamp, no need to set up local server.

Dataset

The PokéChamp dataset contains over 3 million competitive Pokémon battles, filtered to 2 million clean battles across various formats and skill levels. It represents the largest collection of real-player Pokémon battles available for research and AI development.

Dataset Features

• Size: 2 million clean battles (1.9M train, 213K test)
• Formats: 37+ competitive formats from Gen 1-9
• Skill Distribution: Battles across all Elo ranges (1000-1800+)
• Time Period: Battles from multiple months (2024-2025)

Dataset Usage

The dataset is available on HuggingFace milkkarten/pokechamp and can be loaded with:

from datasets import load_dataset
from battle_translate import load_filtered_dataset

# Load the entire dataset performed in load_filtered_dataset
# dataset = load_dataset("milkkarten/pokechamp")

# Load filtered dataset with specific criteria
filtered_dataset = load_filtered_dataset(
    min_month="January2025",
    max_month="March2025",
    elo_ranges=["1600-1799", "1800+"],
    split="train",
    gamemode="gen9ou"
)

# Access battle data
example = next(iter(filtered_dataset))
print(f"Battle format: {example['gamemode']}, Elo: {example['elo']}")

We also provide battle_translate.py for converting raw battle logs into training data using the prompts from our paper:

# Example usage of battle_translate.py
python battle_translate.py \
  --output data/gen9ou_high_elo.json \
  --limit 5000 \
  --gamemode gen9ou \
  --elo_ranges 1600-1799 1800+ \
  --min_month January2025 \
  --max_month March2025 \
  --split train

This script processes battles from the dataset and outputs structured JSON data with instruction-output pairs that can be used for training or evaluating Pokémon battle agents.

Benchmark Puzzles - Coming Soon!

This requires download of our dataset (release TBD). To reproduce the action prediction results:

python evaluate_action_prediction.py

This script will analyze the dataset and output prediction accuracies for player and opponent actions across different Elo ratings.

Acknowledgement

The environment is implemented based on PokeLLMon and Poke Env. This work provides an implementation of the PokéChamp paper:

@article{karten2025pokechamp,
  title={PokéChamp: an Expert-level Minimax Language Agent},
  author={Karten, Seth and Nguyen, Andy Luu and Jin, Chi},
  journal={arXiv preprint arXiv:2503.04094},
  year={2025}
}