slop-detection / README.md
SlavaKemDev's picture
Expose train/test splits with Image features and clean folder names
80cb4c8 verified
|
Raw History Blame Contribute Delete
1.59 kB
metadata
task_categories:
  - image-segmentation
configs:
  - config_name: default
    data_files:
      - split: train
        path: train/metadata.parquet
      - split: test
        path: test/metadata.parquet

Digital Detective AIC 2026 — stage2

Standard train/test splits with individual image and mask files.

from datasets import load_dataset

# Authenticate to this private repository first.
train = load_dataset("SlavaKemDev/slop-detection", split="train")
test = load_dataset("SlavaKemDev/slop-detection", split="test")

Use streaming=True to read examples incrementally. Image columns reference the individual files; the Parquet files contain metadata, not embedded image bytes.

Columns: image, mask, original_image, img_path. Train has 21,275 rows. Test has 1,890 rows; mask and original_image are null because labels are not provided. Test img_path preserves the original competition identifier for submissions.

Files: train/images/, train/originals/, train/masks/, test/images/. Train images use 256 subfolders. train/train.csv points to the new paths. test/test.csv and test/submission.csv preserve the original bytes and identifiers.

Original train archive: https://disk.360.yandex.ru/d/Uyl9N9aiqnJmrA

Original test archive: https://disk.360.yandex.ru/d/Z5iouXhCNlKKww

Downloaded directly through the public Yandex Disk API. Archive SHA256 values match the owner's originals. Image and mask bytes are unchanged.

Test is for final inference only; it must not provide any training or threshold-selection signal.