stevenkhan/pokemon-showdown-battle-sft

🎮 Pokémon Showdown Battle SFT Dataset

500,057 expert battle decisions from high-ELO (1400+) competitive Pokémon Showdown matches, formatted for supervised fine-tuning of language models.

Overview

This dataset converts raw Pokémon Showdown replay logs into instruction-following pairs where:

• User message: Describes the current battle state (active Pokémon, HP, moves, bench, hazards, weather, etc.)
• Assistant message: Recommends the optimal action (which move to use or which Pokémon to switch to) with strategic reasoning

Only the winning player's decisions are included — the model learns from success.

Source Data

Processed from milkkarten/pokechamp, which contains 2.1M+ real Pokémon Showdown battles across all generations.

Filtering Criteria

• ELO ≥ 1400 (competitive-level play)
• Games with a winner (no disconnects/forfeits)
• ≥3 meaningful turns per game
• Winning player's perspective only

Generation Coverage

Generation	Examples
Gen 9 (Scarlet/Violet)	~75%
Gen 8 (Sword/Shield)	~5%
Gen 6 (X/Y)	~3%
Gen 5 (Black/White)	~9%
Gen 3 (Ruby/Sapphire)	~6%
Gen 1 (Red/Blue)	~2%

Format Coverage

Includes OU, RU, Ubers, National Dex, Random Battles, and more across multiple generations.

Data Format

Each example is a list of chat messages:

{
  "messages": [
    {
      "role": "system",
      "content": "You are an expert Pokémon battle strategist for Pokémon Showdown. Analyze the battle state and recommend the optimal action..."
    },
    {
      "role": "user", 
      "content": "Current battle state:\n\nTurn 5 | gen9ou\nField: Weather: sandstorm\n\nYour active Pokémon: Garchomp (75%) (+2 atk)\n  Known moves: earthquake, dragon claw, swords dance\nYour bench: Toxapex (100%), Corviknight (88%)\n\nOpponent's active Pokémon: Dragapult (100%)\nOpponent's other revealed Pokémon: Heatran (50%)\nOpponent's side: stealth rock\n\nWhat should I do this turn?"
    },
    {
      "role": "assistant",
      "content": "**Action: Use earthquake.**\n\nearthquake is super effective against Dragapult. Setting up with swords dance to sweep — this is a safe turn to boost.\n\nGarchomp should use earthquake against Dragapult this turn."
    }
  ]
}

Battle State Features Tracked

• Active Pokémon: Species, HP%, status conditions (PAR, BRN, PSN, etc.)
• Stat boosts: +2 atk, -1 def, etc.
• Known moves: Moves revealed during the battle
• Abilities: When revealed
• Items: When revealed
• Mega Evolution / Terastallization: Tracked
• Bench Pokémon: All team members with HP
• Side conditions: Stealth Rock, Spikes, Toxic Spikes, Screens, etc.
• Field conditions: Weather (rain, sun, sand, hail), Terrain (Electric, Grassy, etc.), Trick Room
• Generation & format: gen1ou, gen9ou, gen8randombattle, etc.

Usage

from datasets import load_dataset

dataset = load_dataset("stevenkhan/pokemon-showdown-battle-sft", split="train")
print(f"Examples: {len(dataset)}")
print(dataset[0])

For SFT Training

from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model="meta-llama/Llama-3.1-8B-Instruct",
    train_dataset=dataset,
    args=SFTConfig(
        max_length=2048,
        assistant_only_loss=True,
    ),
)
trainer.train()

Processing Pipeline

milkkarten/pokechamp (raw Showdown protocol logs)
  ↓ Filter ELO ≥ 1400, games with winner
  ↓ Parse pipe-delimited protocol line by line
  ↓ Track full battle state per turn (both players)
  ↓ Extract winning player's actions only  
  ↓ Generate type-chart-aware strategic reasoning
  ↓ Format as system/user/assistant chat messages
  → 500,057 training examples

Related Resources

• Model: stevenkhan/pokemon-showdown-strategist
• Source data: milkkarten/pokechamp
• PokéChamp paper: arXiv:2503.04094
• Metamon paper: arXiv:2504.04395