OmniCoT Panoramic CoT QA (ImageFolder)

This dataset contains omnidirectional/panoramic images paired with multi-type QA samples, including chain-of-thought (CoT) rationales.

Data format

This repository follows the ImageFolder format with metadata.jsonl for each split:

train/ images/... metadata.jsonl validation/ images/... metadata.jsonl test/ images/... metadata.jsonl real/ images/... metadata.jsonl

Each line in metadata.jsonl is one QA sample. Multiple QA samples may reference the same image via file_name.

Splits

• train: 14,385 QA samples over 2,800 unique images
• validation: 3,060 QA samples over 600 unique images
• test: 3,115 QA samples over 600 unique images
• real: 1,073 QA samples over 200 referenced real-world images

Images are disjoint across splits (no leakage).

Fields

• image (generated by ImageFolder): the panoramic image
• file_name: relative path to the image within the split folder (e.g., images/xxx.png)
• scene_id: scene identifier
• qa_id: unique QA sample id
• type: question type (e.g., viewpoint_transform_identify, ...)
• subtype: subtype label (e.g., A1)
• question: question text
• answer: answer text
• cot: chain-of-thought steps (list of strings; may be empty for real-world samples)
• random_objects: list of objects used for randomization (optional)