training data link: https://contest.hydro.org.cn/T...

I can’t produce a valid submission yet because the supplied training-data archive could not be downloaded from this environment. The PDF alone does not contain the test point_id values or data needed to generate the per-cell CSV files. It also states that the hidden test labels cannot be obtained through non-public channels, so a mathematically maximum score cannot be guaranteed without violating the rules. (Challenge.26UCupSummer.ver.final.pdf)

Please upload TrainingData.26UCupSummer.zip directly here. I’ll then train and validate a model, generate every required <cell-number>.csv, check the IDs and numeric predictions, and package them directly inside a submission-ready output.zip.

You do not need to upload the full 18.8 GB archive. The large size is mainly from the .ply point clouds and visualization images.

Create a compact metadata archive containing only:

• ep.json
• train_signal.csv
• test_points.csv
• sample_submission.csv

Run this Python script in the directory containing the extracted dataset:

python
from pathlib import Path
import zipfile

# Change this to the extracted dataset directory.
dataset_root = Path("TrainingData.26UCupSummer")
output_path = Path("compact_dataset.zip")

required_names = {
    "ep.json",
    "train_signal.csv",
    "test_points.csv",
    "sample_submission.csv",
}

with zipfile.ZipFile(
    output_path,
    mode="w",
    compression=zipfile.ZIP_DEFLATED,
    compresslevel=9,
) as archive:
    count = 0

    for path in dataset_root.rglob("*"):
        if path.is_file() and path.name in required_names:
            relative_path = path.relative_to(dataset_root)
            archive.write(path, relative_path)
            count += 1

print(f"Created {output_path.resolve()}")
print(f"Included {count} files")
print(f"Size: {output_path.stat().st_size / 1024**2:.2f} MB")

Upload compact_dataset.zip here. It will probably be far below 200 MB. From it, I can construct a strong geometry-, antenna-, frequency-, and cross-cell-based submission and return the correctly formatted output.zip.

The point clouds can later be incorporated through compact locally extracted obstruction features rather than uploading all 18.8 GB.

I cannot honestly guarantee the absolute maximum score: the test RSRP labels are hidden, and obtaining or reconstructing them through non-public means is explicitly prohibited. The required output consists of per-cell CSV files placed directly inside output.zip. (Challenge.26UCupSummer.ver.final.pdf)

I tried the exact file.kiwi URL, but it currently returns 404. It is also missing the required decryption fragment. A valid file.kiwi share address has this form:

text
https://file.kiwi/51969dfc#LONG_SECRET_KEY

The portion after # is the only decryption key; without it, the archive cannot be recovered or downloaded. (Invalid URL)

I inspected the attached output.zip: it is structurally valid and contains 19 CSV files with 501,433 predictions. The public scoreboard currently shows PhoenixKiller at 4,077,555, while the leader has 4,501,194, so blindly altering these predictions is unlikely to close the gap. (UCup Challenge by Hydro)

Copy the complete URL from the browser address bar, including everything after #, and paste it here inside backticks. Create a fresh share if the original expired—the free window for an 18.8 GB file is limited. Once the working link is available, I can process the training data and produce a properly validated submission; the absolute maximum cannot be guaranteed because the test labels are hidden. (Challenge.26UCupSummer.ver.final.pdf)