VeraIsHere/sokoban_processed

Sokoban Processed Dataset

This folder contains a processed version of the Sokoban dataset derived from https://huggingface.co/datasets/Xiaofeng77/sokoban. The data is packaged for fast local loading and visual inspection.

• —Files: train.parquet (3,982 rows) and test.parquet (1,602 rows); corresponding PNGs live in images/.
• —Columns:
• —data_source: Source split/config from the original Hugging Face dataset (e.g., sokoban_6x6_1horizon).
• —prompt: Chat-style, multi-turn prompt used to elicit a Sokoban solution.
• —ability: Heuristic/solver used when creating the sample (e.g., bfs).
• —reward_model: Small dict with evaluation info such as ground_truth action indices and style.
• —extra_info: Environment metadata (grid size, initial grid layout, box/player start positions, max steps, render mode).
• —id: Unique example identifier.
• —images: Array with a single base64-encoded PNG string for the board rendering.
• —image_path: Absolute path to the saved PNG inside images/.

Example loading snippet:

import pandas as pd

df = pd.read_parquet("train.parquet")
sample = df.iloc[0]
image_bytes = sample["images"][0].split(",")[1].encode()