GR-IQA

GR-IQA provides image-quality assessment conversations and images for supervised fine-tuning of Zoom-IQA.

Contents

• sft_restoration_v2.json: conversations.
• koniq/images/: original images.
• koniq/crops/: crop images used in follow-up turns.
• dataset_info.json: LLaMA-Factory dataset configuration.

Usage

Records use the ShareGPT format with messages and images fields. Image paths are relative to this repository's root. Crop coordinates are normalized by image width and height.

For LLaMA-Factory, place the data in your data directory, merge dataset_info.json into its dataset configuration, and select sft_restoration_v2.

Source

Images originate from KonIQ-10k/YFCC100M and retain their original licenses. Crops are derivatives of those images. Annotations are provided for research use.