# AgriGround Annotations

AgriGround contains 11,421,148 Stage 4 instruction-tuning records across 503,919 source images and 14 tasks.

| Split | Source images | Annotation records |
|---|---:|---:|
| Train | 401,234 | 9,095,320 |
| Test | 102,685 | 2,325,828 |
| **Total** | **503,919** | **11,421,148** |

## Layout

The `train/` and `test/` directories are grouped by source family and dataset. Each leaf directory contains a gzip-compressed `stage4.jsonl.gz` file. `manifest.json` provides row counts, task counts, compressed sizes, and SHA-256 checksums for every file. `validation.json` records the full-release integrity and schema validation result.

## Record Format

Each line is one JSON object with these common fields:

- `image_path`: path of the image in its source dataset
- `task_type`: machine-readable task identifier
- `question_type`: human-readable task identifier
- `question`: model instruction or question
- `answer`: target response

Task-dependent fields include COCO-style RLE masks, normalized `[x, y, width, height]` bounding boxes, object counts, grounded phrases, regions, and multi-turn conversations. The 14 task definitions and exact distribution are listed in [`docs/TASKS.md`](../docs/TASKS.md).

## Loading

```python
import gzip
import json

with gzip.open("annotations/train/classification/banana_leaf_disease_classification/stage4.jsonl.gz", "rt", encoding="utf-8") as stream:
    for line in stream:
        record = json.loads(line)
        print(record["task_type"], record["question"], record["answer"])
        break
```

Source images are not included. Resolve each `image_path` against the corresponding source dataset and follow that dataset's original license and usage terms.