DATA_SCHEMA.md

# UltraVR QA-Only Data Schema

UltraVR annotations are stored in JSONL format. Each line is one QA sample.

This trial version is QA-only: the original `reasoning_supervision` field has been removed from all questions, and no `gt_cot`, `cot`, or chain-of-thought field is included.

## Required fields

```json
{
  "sample_id": "string, unique UltraVR QA sample id",
  "domain": "RS | CCTV | AD",
  "source_dataset": "DOTA-v1.5 | PANDA | MVTec LOCO AD",
  "image_path": "string",
  "image_width": "integer or null",
  "image_height": "integer or null",
  "question_id": "string or null",
  "question": "string",
  "question_type": "string or null",
  "options": "object, usually A/B/C/D for multiple choice; may be empty for open-answer AD questions",
  "answer": "string",
  "license_note": "string or null"
}
```

## Field notes

- `sample_id` is unique across the release and follows `ultravr_<domain>_<six_digit_id>`.
- `image_path` is the only image locator in each JSONL record. For AD it points to a local repository file under `data/images/ad/`. For RS and CCTV it points to the corresponding source-dataset image path.
- `question_id` preserves the original question identifier from the constructed QA files.
- `question_type` preserves the original category, such as `comparison`, `logical_verification`, or `relational_inference`.
- `options` keeps the original answer options. Some AD anomaly-localization questions are open-answer and therefore have `{}`.
- `answer` keeps the original answer string. For multiple-choice questions it is usually `A`, `B`, `C`, or `D`; for open-answer AD questions it can be a coordinate such as `(5, 10)`.
- `reasoning_supervision` is intentionally omitted.

## Domain image policy

| Domain | `image_path` meaning |
|---|---|
| RS | Source DOTA path, e.g. `DOTA-v1.5_val/P0019.png`. |
| CCTV | Source PANDA path, e.g. `PANDA/image_train/01_University_Canteen/IMG_01_01.jpg`. |
| AD | Local repository path, e.g. `data/images/ad/ad_0001.jpg`. |

## Mapping files

RS and CCTV images are not redistributed. Use these files to map QA samples to source images:

- `data/images/rs/mapping.csv`
- `data/images/cctv/mapping.csv`

All paths use POSIX-style `/` separators and are relative unless they are official source URLs.