datasets:
- jakegrigsby/metamon-synthetic
- jakegrigsby/metamon-parsed-replays
license: apache-2.0
pipeline_tag: reinforcement-learning
library_name: amago
tags:
- pokemon
- game-ai
- offline-rl
- transformers
Checkpoints from Metamon (v1) training runs.
Metamon enables reinforcement learning (RL) research on Pokémon Showdown by providing:
- A standardized suite of teams and opponents for evaluation.
- A large dataset of RL trajectories "reconstructed" from real human battles.
- Starting points for training imitation learning (IL) and RL policies.
Metamon is the codebase behind "Human-Level Competitive Pokémon via Scalable Offline RL and Transformers" (RLC, 2025). Please check out our project website for an overview of our results. This README documents the dataset, pretrained models, training, and evaluation details to help you get battling!
Paper: Human-Level Competitive Pokémon via Scalable Offline Reinforcement Learning with Transformers Project Website: https://metamon.tech Code: GitHub Repository
Usage
Pretrained models can run without research GPUs, but you will need to install amago, which is an RL codebase by the same authors. Follow installation instructions here.
Load and run pretrained models with metamon.rl.eval_pretrained. For example, to run the default checkpoint of the SyntheticRLV2 model for 100 battles against a set of heuristic baselines:
python -m metamon.rl.eval_pretrained --agent SyntheticRLV2 --gens 1 --formats ou --n_challenges 100 --eval_type heuristic
To battle against other models or humans online (via a local Showdown server):
python -m metamon.rl.eval_pretrained --agent SyntheticRLV2 --gens 1 --formats ou --n_challenges 50 --eval_type ladder --username <pick unique username> --team_set competitive
For more details on models and usage, please refer to the project's GitHub repository.