LoRA adapter for Champion tier play. Trained with battle-outcome GRPO on gen9randombattle data (reward from wins vs heuristics and baselines).

Stack on Qwen3-4B or load after the full SFT model for best ladder strength.

Quick start

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "Qwen/Qwen3-4B",
    max_seq_length=2048,
    load_in_4bit=False,
)
model.load_adapter("GoldenGrapeGentleman1/pokemon-showdown-agent-battle-grpo", adapter_name="grpo")
model.set_adapter("grpo")

Use in this tutorial

Step 0:

python

os.environ["POKEMON_AGENT_TIER"] = "champion"
os.environ["POKEMON_HF_BATTLE_GRPO"] = "GoldenGrapeGentleman1/pokemon-showdown-agent-battle-grpo"

Then run Steps 13 → 14 → 15.