homerquan/boardgamebench-answer-only

BoardGameBench Answer-Only Reasoning Dataset

This dataset contains 1,282,766 board-game reasoning examples generated from BoardGameBench, a benchmark and data-generation project for evaluating language models on structured board-game decision making.

Each row asks a model to inspect a legal board position and return the best move. The format is intentionally simple:

id,prompt,answer

The answer field is the target move label, such as C4, f6, 11,7, or e2-d3. This makes the dataset easy to use for supervised fine-tuning, exact-match evaluation, or as a seed corpus for preference-learning experiments.

Why This Dataset Exists

Board games are a practical way to study reasoning because they combine:

• —clear rules and legal-action constraints,
• —tactical and strategic search,
• —verifiable outcomes from game engines,
• —compact prompts with exact target answers.

The examples were generated from BoardGameBench game states and engine-guided move selection. They are designed to help language models learn to parse board state, respect legal moves, and choose stronger actions across several classic games.

Included Games

The dataset includes positions from BoardGameBench games such as:

• —Connect Four
• —Gomoku 19x19
• —Othello 6x6
• —Othello 8x8
• —Hex 7x7
• —Breakthrough 6x6
• —Dots and Boxes 3x3

Files

• —answer_only.csv: full CSV dataset with id, prompt, and answer.
• —jsonl/: JSONL shards with the same fields, split into 5,000-row shards.
• —manifest.json: source manifest and row count.

Example

{
  "id": "00004707bc9b20f053004c61",
  "prompt": "Below is an instruction that describes a board-game reasoning task...",
  "answer": "11,7"
}

Loading

from datasets import load_dataset

ds = load_dataset("homerquan/boardgamebench-answer-only", data_files="jsonl/*.jsonl")
print(ds["train"][0])

Suggested Uses

• —Supervised fine-tuning of language models to produce exact board-game moves.
• —Evaluation of legal move selection and tactical board-state understanding.
• —Seed data for DPO or other preference-training pipelines.
• —Research into compact, verifiable reasoning tasks.

Source Project

This dataset is built from the ideas and tooling in:

https://github.com/homerquan/BoardGameBench

If you use this dataset, please also check out the BoardGameBench project, where the benchmark and game-generation ideas are developed.

Notes

The dataset focuses on answer-only supervision. It does not include hidden chain-of-thought. Some source generation pipelines also produced preference and verifier-style rows, but this release intentionally publishes the clean id,prompt,answer version for broad compatibility.