# Publishing the Dataset to HuggingFace Hub

## Files

| File | Purpose |
|---|---|
| `build_dataset.py` | Converts `CzechHistoricalNewspapers.json` into two HuggingFace Dataset configs |
| `dataset_card.md` | Hub README — copy as `README.md` in the Hub repo before pushing |

---

## Dataset configs produced

| Config | Rows | Granularity | Best for |
|---|---|---|---|
| `pages` | 30 | One row per newspaper page | Running clustering experiments (texts list + labels list per page) |
| `annotations` | 881 | One row per text annotation | Cross-page filtering, per-annotation analysis |

---

## Workflow

### 1. Install dependencies

```bash
pip install datasets huggingface_hub
```

### 2. Build locally (no Hub account needed)

```bash
python build_dataset.py
# Output saved to ./hf_dataset/pages/ and ./hf_dataset/annotations/
```

### 3. Log in and push to Hub

```bash
huggingface-cli login
python build_dataset.py --push your-hf-username
# Publishes as: your-hf-username/czech-historical-newspapers
```

### 4. Add the dataset card

Copy `dataset_card.md` as `README.md` in the Hub repo (via the Hub UI or `huggingface_hub` API).

---

## Loading the dataset from anywhere

```python
from datasets import load_dataset

# Page-level (for clustering experiments)
ds = load_dataset("your-username/czech-historical-newspapers", "pages")
row = ds["train"][0]
texts  = row["texts"]           # list of OCR strings for the page
labels = row["cluster_labels"]  # ground-truth article assignment per text

# Annotation-level (for per-region analysis)
ds = load_dataset("your-username/czech-historical-newspapers", "annotations")
page = ds["train"].filter(lambda x: x["image_name"] == "Brnenske_48_D.jpg")
```

---

## Key design decisions

- **Two configs** instead of one flat table: the page-level view avoids reconstructing the grouping logic in every experiment notebook.
- **Cluster label alignment**: the script iterates annotations in the original JSON order (same as the notebook), so `cluster_labels[i]` always corresponds to `texts[i]`.
- **Filtered categories**: region types 1 (SeparatorRegion), 3 (Caption), 6 (ImageRegion), and 7 (Article) are excluded — they carry no OCR text and are not part of the cluster arrays.
- **Images not included**: the original JPEGs are not in the JSON. The dataset is text-only; `image_name` serves as a reference key.