Fine-tuned Qwen3-4B on 2.3M turn-based battle game replay logs. The model reads raw battle protocol lines and outputs the next action as move … or switch ….

Use case: Competitive tier agent for gen9randombattle — load directly for inference or as the base for battle-oriented GRPO.

Quick start

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="GoldenGrapeGentleman1/battle game-showdown-agent-full-sft",
    max_seq_length=2048,
    load_in_4bit=False,
)

Training

Setting	Value
Base model	Qwen/Qwen3-4B
Data	Raw Showdown replay logs (~2.3M samples)
Method	Full SFT (merged weights)
Hardware	AMD Instinct MI300X, ROCm, bfloat16

Tutorial

End-to-end ROCm notebook: battle game LLM Agent with Unsloth — set POKEMON_AGENT_TIER=competitive and POKEMON_HF_FULL_SFT=GoldenGrapeGentleman1/battle game-showdown-agent-full-sft.

License

Apache-2.0 (base model Qwen3-4B).

GoldenGrapeGentleman1/battle-game-agent-full-sft

turn-based battle game Agent — Full SFT

Quick start

Training

Tutorial

License

GoldenGrapeGentleman1/
battle-game-agent-full-sft