MindGames Challenge

MindGames reproduction — 4 games, top-3 approaches, Qwen3-8B, 5 different configurations.

Folder	What
eunbi/	Shared LoRA train + eval
mgc2025_sft/	MGC2025 → LoRA
in2ai/	1st — GRPO
stars/	2nd — Ollama ReAct
rlgaming/	3rd — filtered SFT

Setup

cp .env.example .env && conda activate maf
bash scripts/install-conda-hook.sh   # one-time
pip install -r requirements.txt

Replace blotto with mafia / ipd / codenames. Set run name: RUN_ID=my_run bash ...

---

Command

mgc2025_sft/

	Command
train	MGC_SMOKE_CONVERT_ONLY=1 MGC_SMOKE_GAME=blotto bash mgc2025_sft/run_smoke.sh (data) then python eunbi/lora/train_lora.py --train_file mgc2025_sft/data/runs/$RUN_ID/blotto_train.jsonl --run_id $RUN_ID --load_in_4bit --epochs 1
eval	python mgc2025_sft/evaluate.py --game blotto --model_dir eunbi/lora/runs/$RUN_ID --test_file mgc2025_sft/data/runs/$RUN_ID/blotto_test.jsonl --n 100
train+eval	bash mgc2025_sft/run_blotto.sh
all games	bash mgc2025_sft/run_all.sh
smoke	bash mgc2025_sft/run_smoke.sh

in2ai/

	Command
train	bash in2ai/run_blotto.sh
eval	—
train+eval	bash in2ai/run_blotto.sh
all games	bash in2ai/run_all.sh
smoke	RUN_ID=fp_smoke N_EPISODES=4 bash in2ai/run_blotto.sh

stars/

	Command
train	—
eval	bash stars/run_blotto.sh
train+eval	—
all games	bash stars/run_all.sh
dry-run	STARS_DRY_RUN=1 bash stars/run_blotto.sh

rlgaming/

	Command
train	RLG_GAME=blotto bash rlgaming/run_data.sh (data only)
eval	python rlgaming/eval/evaluate.py --game blotto --model_dir rlgaming/runs/$RUN_ID/lora --test_file rlgaming/runs/$RUN_ID/blotto_test.jsonl --n 100
train+eval	bash rlgaming/run_blotto.sh
all games	bash rlgaming/run_all.sh
smoke	bash rlgaming/run_smoke.sh

eunbi/ (Implement with Lora)

LoRA: r=16, alpha=32, dropout=0.05, 4-bit, Qwen3-8B, targets all 7 linear layers (q/k/v/o_proj, gate/up/down_proj). Same in in2ai GRPO.

	Command
train	python eunbi/lora/train_lora.py --train_file <jsonl> --run_id $RUN_ID --load_in_4bit --epochs 1
eval	python eunbi/eval/evaluate_blotto.py --model_dir eunbi/lora/runs/$RUN_ID --run_id $RUN_ID --n_games 30
train+eval	bash eunbi/run_blotto.sh
all games	bash eunbi/run_all.sh
smoke	bash eunbi/run_smoke.sh

---

Outputs

eunbi/lora/runs/<RUN_ID>/ · eunbi/eval/runs/<RUN_ID>/ · eunbi/data/runs/<RUN_ID>/ · mgc2025_sft/data/runs/<RUN_ID>/ · in2ai/runs/<RUN_ID>/ · stars/runs/<RUN_ID>/ · rlgaming/runs/<RUN_ID>/

Env (.env)

MAF_BASE_MODEL=Qwen/Qwen3-8B · MAF_LORA_4BIT=1 · STARS_OLLAMA_HOST=http://127.0.0.1:11434