homerquan/nemotron-boardgame-answer-lora-b4-safe-2000

Nemotron Boardgame Answer LoRA B4 Safe 2000

Overview

This is a specialist PEFT/LoRA adapter for playing board games with language-model prompts. It is intended only for board-game reasoning, move selection, and structured game-answer generation. It is not a general assistant, chat model, coding model, search model, or safety-critical decision system.

In the supplied BoardGameBench run, this adapter reaches a 271.9 BRI over 22 games, outperforming the listed top-line open-model baselines for this board-game benchmark snapshot. The adapter is designed for compact deterministic board games where the model must infer the board state, choose legal and strategically strong moves, and answer in a stable format.

Author: Homer Quan
Project: BoardGameBench

What This Is

nemotron-boardgame-answer-lora-b4-safe-2000 is a LoRA adapter trained for board-game answering on top of nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16. The repository contains adapter weights and tokenizer assets for PEFT loading.

The adapter is narrow by design:

• It should be used for board-game play and board-game benchmark evaluation.
• It should be prompted with explicit game state, rules, legal moves when available, and the expected answer format.
• It should not be treated as a general-purpose instruction model.

Benchmark Snapshot

The table below summarizes the provided BoardGameBench comparison. Lower raw loss is not the headline here; the main comparison metric is BRI.

Model	Games	Wins	Losses	Draws	Forfeits	Raw Score	Normalized	BRI
nemotron_nano_b4_safe_2000_adapter	22	1	21	0	3	5.37/22	24.42	271.9
Nemotron 3 33B (Ollama)	70	2	68	0	38	4.91/70	7.02	118.0
Nemotron 3 Super (Ollama)	70	0	70	0	36	3.02/70	4.31	97.3
Gemma 4 8B (Ollama)	70	0	70	0	60	0.71/70	1.01	71.2

These results are from the supplied BoardGameBench reports and should be read as benchmark-specific performance, not a general intelligence claim. Re-run the benchmark with your own hardware, prompt format, base-model revision, and evaluation harness version before making production comparisons.

Intended Use

Use this adapter when you want a model that is deliberately biased toward board-game play:

• choosing moves in compact deterministic board games
• answering board-game benchmark prompts
• comparing game-play behavior across open models
• experimenting with prompt formats for BoardGameBench

Out of scope:

• general chat or instruction following
• factual QA, medical, legal, financial, or safety advice
• autonomous real-world decisions
• games or tasks where the rules are not included or are ambiguous

Loading Example

This repository is a PEFT adapter, not a standalone base model. Load it with the Nemotron 3 Nano base model used for training.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
adapter_id = "YOUR_HF_USERNAME/nemotron-boardgame-answer-lora-b4-safe-2000"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
    torch_dtype="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

messages = [
    {
        "role": "system",
        "content": "You are a board-game playing model. Answer only with the requested move format.",
    },
    {
        "role": "user",
        "content": "Game: ...\nBoard: ...\nLegal moves: ...\nChoose the best move.",
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Prompting Guidance

For best results, keep prompts concrete and game-like:

• name the game
• include the full board state
• include current player and turn context
• include legal moves when available
• ask for a single move or a compact structured answer
• avoid open-ended chat instructions

Example prompt shape:

Game: <game name>
Current player: <player>
Board:
<state>
Legal moves:
<moves>

Return the best move only.

Training Details

• Adapter type: LoRA
• PEFT task type: causal language modeling
• Rank: 8
• LoRA alpha: 16
• LoRA dropout: 0.0
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• PEFT version: 0.19.1

The adapter was trained for board-game answer behavior. Training data details are not bundled in this adapter repository.

Limitations

• The adapter depends on the compatible base model. Loading it on an incompatible checkpoint can produce poor or invalid outputs.
• It may still choose illegal moves when the state is underspecified or legal moves are omitted.
• It is optimized for board-game benchmark behavior, not for broad reasoning tasks.
• Benchmark results can vary with prompt format, decoding settings, base model revision, and evaluation harness version.

Citation

If you use this adapter in evaluations or derivative work, please cite the BoardGameBench project:

@software{quan_boardgamebench_2026,
  author = {Quan, Homer},
  title = {BoardGameBench},
  year = {2026},
  url = {https://github.com/homerquan/BoardGameBench}
}

Author

Author: Homer Quan

Framework Versions

• PEFT 0.19.1