Download Player-1-Data: 6,044 trajectories from 172 games, published as JSON and JPEGs. Follow the dataset card to prepare the training format:
dataset/
data/
manifest.json
train.parquet
validation.parquet
images/trajectories/<trajectory-id>/<frame>.jpg
Set PLAYER1_DATASET_ROOT to this directory for SFT. The published conversion has an empty validation split.
Install dependencies:
conda run -n verl pip install -r requirements/data.txtUse the Dataset class to read trajectories with reconstructed visual-action history:
from pathlib import Path
from player1.datasets.dataset import Dataset
data_root = Path("/path/to/Player-1-Data")
dataset = Dataset(
data_root,
split="train",
recent_turns=2, # Turns with full screenshots
old_turns=2, # Older turns retaining actions only
)
try:
example = dataset[0]
print("Instruction:", example["messages"][0]["content"])
print("Target action:", example["messages"][-1]["content"])
print(f"Supervised message count: {sum(example['message_loss_mask'])}")
finally:
dataset.close()Each Parquet row group is one trajectory. Rows store the current screenshot, instructions, reasoning, and canonical action; data/manifest.json records provenance and splits. Keep data/ and images/ together.
read_dataset.py reconstructs recent visual and older action-only history. See Data production to create datasets or open their Viewer:
python -m player1_data view /path/to/Player-1-Data --port 8765