---
dataset_info:
  features:
  - name: images
    list:
      image:
        decode: false
  - name: image_group_id
    dtype: string
  - name: qa
    list:
    - name: inputs
      dtype: string
    - name: labels
      dtype: string
    - name: metadata
      dtype: string
  splits:
  - name: test
    num_bytes: 6166925
    num_examples: 500
  - name: train
    num_bytes: 307575696
    num_examples: 30000
  download_size: 314840221
  dataset_size: 313742621
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
  - split: train
    path: data/train-*
license: apache-2.0
source_datasets:
- Valen-Team/Valen-Eval-Game
---

# valen-game

Valen Sokoban: native optimal-action sets as action-conditioned yes/no decisions.

All four native directions are supervised, including every equally optimal
action. Original English/Chinese rules, train/test layout partitions, image
bytes, native probabilities and solver provenance are retained. Hidden states
and solutions stay in metadata. Four QAs share one screenshot. Evaluation is
the source's selected single-step subset, not an unbiased full-game benchmark.
Original source declares Apache 2.0. No validation split is fabricated.

Original data: [Valen-Team/Valen-Eval-Game](https://huggingface.co/datasets/Valen-Team/Valen-Eval-Game). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/games.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/games.py#L129).

[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.