homerquan/boardgamebench-answer-dpo

BoardGameBench Answer DPO Dataset

This dataset contains the reviewed preference examples used for the DPO stage of the nemotron-boardgame-answer-lora-b4-safe-final adapter.

It is a compact pilot set of 10 BoardGameBench preference rows. Each row presents the same board-game decision prompt with a preferred answer and a plausible rejected answer. The preferred answer is selected from engine-guided move comparisons and includes the exact move label.

Format

The main training file is:

• dpo_train.jsonl

Each JSONL row includes:

• id: stable example id.
• prompt: board-game decision prompt.
• chosen: preferred response text.
• rejected: rejected response text.
• chosen_move: exact preferred move label.
• rejected_move: exact rejected move label.
• score_margin: engine/search margin between the two moves.
• metadata_json: serialized source metadata.

Example

{
  "id": "62b39b7d6772a5c5b8ace2b4",
  "prompt": "Below is an instruction that describes a board-game reasoning task...",
  "chosen_move": "C4",
  "rejected_move": "C7",
  "score_margin": 3.0
}

Loading

from datasets import load_dataset

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

Training Lineage

This DPO set was used after the SFT stage and before GRPO:

1. SFT on homerquan/boardgamebench-answer-sft.
2. DPO on this reviewed preference dataset.
3. GRPO on homerquan/boardgamebench-answer-grpo.

The DPO set is intentionally small because it was a reviewed pilot pass, not the large SFT corpus. It was used to nudge the SFT adapter toward clearer preference behavior before the GRPO reward-map stage.

Source Project

The prompts, legal moves, and engine-guided comparisons come from:

https://github.com/homerquan/BoardGameBench

Notes

Rows may contain concise synthetic rationale text in chosen and rejected; the exact move labels are also exposed separately for easier filtering and analysis.