README.md

metadata

pretty_name: Player-1-Data
language:
  - en
task_categories:
  - image-text-to-text
tags:
  - gaming
  - imitation-learning
  - keyboard-mouse-control
size_categories:
  - 1K<n<10K
license: other
license_name: source-dependent
license_link: https://huggingface.co/datasets/yyyang/Player-1-Data/blob/main/SOURCES.md
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.json
dataset_info:
  features:
    - name: id
      dtype: string
    - name: game
      dtype: string
    - name: mode
      dtype: string
    - name: source
      struct:
        - name: recording_id
          dtype: string
        - name: action_clock
          dtype: string
        - name: action_interval_ns
          list: int64
        - name: observation_span_ns
          list: int64
        - name: raw_operation_clock
          dtype: string
    - name: issues
      list: string
    - name: initial_messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
    - name: observation_prompt
      dtype: string
    - name: turns
      list:
        - name: image
          dtype: string
        - name: reasoning
          dtype: string
        - name: action
          dtype: string
        - name: source
          struct:
            - name: action_interval_ns
              list: int64
            - name: observation_time_ns
              dtype: int64
            - name: frame_index
              dtype: int64
            - name: raw_operation_interval_ns
              list: int64

Player-1-Data

Player-1-Data is a collection of gameplay demonstrations for training game-control agents. It contains 6,044 trajectories and 295,207 turns, including gameplay and in-game GUI interactions. Each turn pairs a screenshot with reasoning and a keyboard/mouse action.

The dataset is part of Player-1: Learning to Play from Synthetic Experience; see the paper for the overall method.

Contents

File	Description
data/<trajectory_id>.json	One trajectory per file
images/<trajectory_id>/	Screenshots referenced by the trajectory
recordings.jsonl	Source recording and session references
SOURCES.md	Source links and usage terms
to_parquet.py	Converter to the SFT training format

Each trajectory has a id, game, mode, source information, initial messages and ordered turns. A turn contains image, reasoning, action and source timing information. Image paths are relative to images/. Turns are stored individually; a reader can assemble history as needed.

Sources

The trajectories come from five recording sources:

Source	Trajectories
GameHorizon Suite	1,761
D2E	1,005
Gaming500	985
Pixel2Play	1,140
VPT	1,153

recordings.jsonl connects trajectories to their original sessions and filenames. Original recordings are not included. The action and video timestamps may use different clocks; use the declared clock and frame index when relating them. See SOURCES.md for source links and license terms; source and third-party game-content terms apply.

Read a trajectory

import json
from pathlib import Path

root = Path("/path/to/Player-1-Data")
trajectory = json.loads(next((root / "data").glob("*.json")).read_text(encoding="utf-8"))
turn = trajectory["turns"][0]
image_path = root / "images" / turn["image"]
print(trajectory["game"], image_path, turn["reasoning"], turn["action"])

Training format

Convert the file tree into training Parquet and images using:

python /path/to/Player-1-Data/to_parquet.py --output /path/to/output

The output contains one Parquet row group per trajectory.

Citation

@misc{ouyang2026player1learningplaysynthetic,
      title={Player-1: Learning to Play from Synthetic Experience}, 
      author={Mingyu Ouyang and Yiran Wang and Siyuan Hu and Xingyilang Yin and Huiqiang Sun and Guangzhi Wang and Junfu Pu and Cheng Cheng and Honghao Chen and Ying Shan and Mike Zheng Shou and Hwee Tou Ng},
      year={2026},
      eprint={2610.20584},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.20584}, 
}