# Preview samples

**Quick human-readable look at the dataset in its original schema.**

> Parquet under `data/` is the canonical format for benchmarking and Hugging Face loading. This folder exists purely so reviewers and other researchers can skim a handful of real rows without writing any code or installing  `datasets`.

## Contents

| File | Count | Notes |
|---|---|---|
| `preview.jsonl` | **20** rows | Stratified sample from `data/canonical.parquet`, in the source-native JSON-per-line schema. Fields are identical to those documented in the root `README.md`. |
| `vce_plus_preview.jsonl` | **20** rows | VCE+ 7-dimensional structured visual-evidence extractions **aligned by `instance_id`** to `preview.jsonl`, illustrating the auxiliary schema used by text-only baselines. |
| `images/` | 31 files | Every image referenced by `preview.jsonl.images[*].local_path`. Filenames follow `issue_<issue_id>_<idx>.{png,jpg,...}`. |

## Quick inspection

```bash
# Pretty-print the first canonical row
head -1 preview.jsonl | python3 -m json.tool

# All gold files across the 20 samples
python3 -c "import json; [print(r['instance_id'], '→', r['edit_files']) \
            for r in map(json.loads, open('preview.jsonl'))]"

# VCE+ extraction for the same instance — side-by-side schema comparison
python3 - <<'PY'
import json
canon = {json.loads(l)["instance_id"]: json.loads(l) for l in open("preview.jsonl")}
vce   = {json.loads(l)["instance_id"]: json.loads(l) for l in open("vce_plus_preview.jsonl")}
iid = next(iter(canon))
print("instance:", iid)
print("  issue title   :", canon[iid]["issue_title"])
print("  image category:", canon[iid]["image_category"])
print("  VCE+ record[0]:", json.dumps(vce[iid]["records"][0], indent=2, ensure_ascii=False))
PY

# Open the image for the first sample
python3 -c "import json; r=json.loads(open('preview.jsonl').readline()); \
            print('first image:', r['images'][0]['local_path'])"
```

## Full dataset

To work with all 652 instances, all 1050 images (embedded as bytes), and both evaluation granularities, use the Parquet configs from the root:

```python
from datasets import load_dataset
ds = load_dataset("<hub-id>/MM-IssueLoc-Bench", name="canonical", split="test")
```

See the root `README.md` and `examples/load_dataset.py` for full workflows.