Mabyduck/CLIC2024-test-human-eval

CLIC 2024 Testset Human Evaluation Dataset

This dataset contains human evaluations of compression quality between images from the CLIC 2024 challenge image track test set. Comparisons were collected on the Mabyduck platform using a 2AFC (two-alternative forced choice) protocol with random sampling of image pairs.

Note: This dataset is not the official CLIC 2024 data used to compute challenge rankings. It reproduces the CLIC 2024 evaluation environment for testing and research purposes only.

It is split into two subsets:

• Screened: 1920 evaluations from verified raters.
• Unscreened: 1921 evaluations from unverified raters.

Dataset Fields

• id: Unique ID for the evaluation.
• rater_id: ID of the human rater.
• session_id: ID of the evaluation session.
• job_id: ID of the job associated with the evaluation.
• ratings: List of individual ratings per image:
  • score: +1 for win, -1 for loss in the 2AFC comparison.
  • meta: Metadata for the stimulus (e.g., crop dimensions).
  • condition: Method used to compress the reference image.
  • stimulus: Stimulus filename or identifier.
  • stimulus_id: Stimulus numeric ID.
• date_updated: Date of the last update.
• is_training: Boolean indicating if this evaluation is part of rater training data.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Mabyduck/CLIC2024-test-human-eval")

# Access splits
screened = dataset["screened"]
unscreened = dataset["unscreened"]

Citation

If you use this dataset, please cite it as:

@misc{Mabyduck_CLIC2024_test_human_eval,
  title        = {CLIC 2024 Test Human Evaluation Dataset},
  author       = {Mabyduck},
  howpublished = {\url{https://huggingface.co/datasets/Mabyduck/CLIC2024-test-human-eval}},
  note         = {Accessed on Hugging Face}
}