yuuu94/PrismWF-ARES-Splits

PrismWF ARES Data Splits

This repository provides the seed-2024 train, validation, and test partitions used by PrismWF for the released ARES closed-world and open-world 2--5-tab benchmarks. Each setting includes the complete partitioned NPZ files, exact split indices, and split metadata.

The partitions are derived from the traffic traces published in the official Zenodo record. Please retain the original dataset citation when using these files.

Protocol

The indices follow the WFlib instance-level protocol:

1. split all trace instances into 90% train-validation and 10% test data;
2. split the train-validation partition into 90% training and 10% validation data;
3. use random seed 2024 for both operations.

This yields an 81/9/10 train/validation/test ratio.

Scenario	Source instances	Train	Validation	Test
Closed-world, 2--5 tabs	58,000 per setting	46,980	5,220	5,800
Open-world, 2--5 tabs	64,000 per setting	51,840	5,760	6,400

Files

Each setting contains:

closed_2tab/
  train.npz
  valid.npz
  test.npz
  split_indices.npz
  split_metadata.json

Each partitioned NPZ contains the trace array X and multi-hot label array y. split_indices.npz stores the corresponding integer arrays train, valid, and test. The same structure is provided for closed_2tab through closed_5tab and open_2tab through open_5tab.

Direct Download

from huggingface_hub import hf_hub_download

train_path = hf_hub_download(
    repo_id="yuuu94/PrismWF-ARES-Splits",
    repo_type="dataset",
    filename="closed_2tab/train.npz",
)

Reconstruct the NPZ Splits

from pathlib import Path

import numpy as np
from huggingface_hub import hf_hub_download

dataset = "closed_2tab"
source_path = Path(f"datasets/{dataset}.npz")
output_dir = Path("datasets") / dataset
output_dir.mkdir(parents=True, exist_ok=True)

index_path = hf_hub_download(
    repo_id="yuuu94/PrismWF-ARES-Splits",
    repo_type="dataset",
    filename=f"{dataset}/split_indices.npz",
)

with np.load(source_path) as source, np.load(index_path) as indices:
    for split in ("train", "valid", "test"):
        selected = indices[split]
        np.savez(
            output_dir / f"{split}.npz",
            X=source["X"][selected],
            y=source["y"][selected],
        )

The source traces and these derived partitions remain subject to the terms and citation requirements of the original release.

Citation

Please cite both the ARES dataset publication and the PrismWF manuscript when using these partitions.