scripts/prepare_huggingface.py

#!/usr/bin/env python3
"""
Prepare UVW 2026 dataset for HuggingFace Hub upload.

UVW 2026: Underthesea Vietnamese Wikipedia Dataset
https://github.com/undertheseanlp/underthesea/issues/896

Uses streaming to handle large datasets efficiently.
Supports direct upload to HuggingFace Hub with --push flag.
"""

import argparse
import json
from datetime import datetime
from pathlib import Path

from tqdm import tqdm

try:
    from datasets import Dataset, DatasetDict, Features, Value
    HF_AVAILABLE = True
except ImportError:
    HF_AVAILABLE = False

try:
    from huggingface_hub import HfApi, login
    HF_HUB_AVAILABLE = True
except ImportError:
    HF_HUB_AVAILABLE = False


SPLITS_DIR = Path(__file__).parent.parent / "data" / "splits"
HF_DIR = Path(__file__).parent.parent / "data" / "huggingface"


def count_lines(path: Path) -> int:
    """Count lines without loading into memory."""
    count = 0
    with open(path, "r", encoding="utf-8") as f:
        for _ in f:
            count += 1
    return count


def iter_jsonl(path: Path):
    """Iterate over JSONL file without loading all into memory."""
    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            yield json.loads(line)


def create_hf_dataset():
    """Create HuggingFace dataset from splits using streaming."""
    if not HF_AVAILABLE:
        print("Cannot create HuggingFace dataset: datasets library not installed")
        print("Install with: uv sync --extra huggingface")
        return

    # Define features (including new wikidata fields)
    features = Features({
        "id": Value("string"),
        "title": Value("string"),
        "content": Value("string"),
        "num_chars": Value("int32"),
        "num_sentences": Value("int32"),
        "quality_score": Value("int32"),
        "wikidata_id": Value("string"),
        "main_category": Value("string"),
    })

    # Load splits using generator for memory efficiency
    splits = {}
    split_counts = {}

    for split_name in ["train", "dev", "test"]:
        jsonl_path = SPLITS_DIR / f"{split_name}.jsonl"
        if not jsonl_path.exists():
            continue

        # Count lines first
        count = count_lines(jsonl_path)
        split_counts[split_name] = count

        # Load data (HuggingFace datasets handles this efficiently)
        print(f"  Loading {split_name}...")
        data = []
        for article in tqdm(iter_jsonl(jsonl_path), total=count, desc=f"    {split_name}"):
            # Ensure all fields exist with defaults
            data.append({
                "id": article.get("id", ""),
                "title": article.get("title", ""),
                "content": article.get("content", ""),
                "num_chars": article.get("num_chars", 0),
                "num_sentences": article.get("num_sentences", 0),
                "quality_score": article.get("quality_score", 0),
                "wikidata_id": article.get("wikidata_id") or "",
                "main_category": article.get("main_category") or "",
            })

        # Rename 'dev' to 'validation' for HuggingFace convention
        hf_split_name = "validation" if split_name == "dev" else split_name
        splits[hf_split_name] = Dataset.from_list(data, features=features)
        print(f"    Loaded {split_name}: {count:,} examples")

        # Free memory
        del data

    if not splits:
        print("No splits found. Please run create_splits.py first.")
        return

    # Create DatasetDict
    dataset = DatasetDict(splits)

    # Save to disk
    HF_DIR.mkdir(parents=True, exist_ok=True)
    print(f"\nSaving dataset to disk...")
    dataset.save_to_disk(HF_DIR / "uvw_2026")
    print(f"  Dataset saved to: {HF_DIR / 'uvw_2026'}")

    # Also save as parquet for easy upload
    parquet_dir = HF_DIR / "uvw_2026_parquet"
    parquet_dir.mkdir(parents=True, exist_ok=True)

    print(f"\nSaving parquet files...")
    for split_name, split_dataset in dataset.items():
        parquet_path = parquet_dir / f"{split_name}.parquet"
        split_dataset.to_parquet(parquet_path)
        print(f"  Saved {parquet_path}")

    return dataset, split_counts


def load_dataset_statistics():
    """Load statistics from processed data files."""
    stats = {
        "total_articles": 0,
        "total_chars": 0,
        "total_sentences": 0,
        "with_wikidata": 0,
        "with_category": 0,
        "categories": {},
        "quality_distribution": {i: 0 for i in range(1, 11)},
    }

    # Try to load from wikidata file (most complete)
    wikidata_path = Path(__file__).parent.parent / "data" / "processed" / "uvw_2026_wikidata.jsonl"
    if not wikidata_path.exists():
        wikidata_path = Path(__file__).parent.parent / "data" / "processed" / "uvw_2026_quality.jsonl"
    if not wikidata_path.exists():
        wikidata_path = Path(__file__).parent.parent / "data" / "processed" / "uvw_2026.jsonl"

    if wikidata_path.exists():
        print(f"  Loading statistics from {wikidata_path.name}...")
        for article in tqdm(iter_jsonl(wikidata_path), desc="    Calculating stats"):
            stats["total_articles"] += 1
            stats["total_chars"] += article.get("num_chars", 0)
            stats["total_sentences"] += article.get("num_sentences", 0)

            if article.get("wikidata_id"):
                stats["with_wikidata"] += 1

            category = article.get("main_category")
            if category:
                stats["with_category"] += 1
                stats["categories"][category] = stats["categories"].get(category, 0) + 1

            quality = article.get("quality_score", 0)
            if 1 <= quality <= 10:
                stats["quality_distribution"][quality] += 1

    return stats


def create_dataset_card(split_counts: dict = None, stats: dict = None):
    """Create README.md for HuggingFace dataset."""
    # Load metadata
    metadata_path = Path(__file__).parent.parent / "data" / "processed" / "metadata.json"
    if metadata_path.exists():
        with open(metadata_path, "r", encoding="utf-8") as f:
            metadata = json.load(f)
    else:
        metadata = {"statistics": {}}

    file_stats = metadata.get("statistics", {})

    # Calculate split counts
    if split_counts:
        total = sum(split_counts.values())
        train_count = split_counts.get("train", 0)
        val_count = split_counts.get("dev", 0)
        test_count = split_counts.get("test", 0)
    else:
        total = file_stats.get("num_articles", 0)
        train_count = int(total * 0.8) if total else "80%"
        val_count = int(total * 0.1) if total else "10%"
        test_count = int(total * 0.1) if total else "10%"

    # Use provided stats or defaults
    if stats and stats["total_articles"] > 0:
        wikidata_pct = (stats["with_wikidata"] / stats["total_articles"] * 100)
        category_pct = (stats["with_category"] / stats["total_articles"] * 100)
        unique_categories = len(stats["categories"])
        avg_chars = stats["total_chars"] // stats["total_articles"]
        avg_sentences = stats["total_sentences"] // stats["total_articles"]

        # Get top categories
        sorted_cats = sorted(stats["categories"].items(), key=lambda x: -x[1])[:10]
        top_categories_table = "\n".join(
            f"| {cat} | {count:,} | {count/stats['total_articles']*100:.1f}% |"
            for cat, count in sorted_cats
        )

        # Quality distribution table
        quality_table = "\n".join(
            f"| {score} | {count:,} | {count/stats['total_articles']*100:.1f}% |"
            for score, count in sorted(stats["quality_distribution"].items())
            if count > 0
        )
    else:
        wikidata_pct = 99.4
        category_pct = 97.0
        unique_categories = 11549
        avg_chars = 2500
        avg_sentences = 30
        top_categories_table = """| đơn vị phân loại (taxon) | 618,281 | 55.3% |
| người (human) | 78,191 | 7.0% |
| xã của Pháp | 35,635 | 3.2% |
| khu định cư | 20,276 | 1.8% |
| tiểu hành tinh | 17,891 | 1.6% |
| xã của Việt Nam | 7,088 | 0.6% |"""
        quality_table = """| 1 | - | - |
| 5 | - | - |
| 10 | - | - |"""

    # Determine size category
    if total == "N/A" or total == 0:
        size_category = "n<1K"
    elif total < 1000:
        size_category = "n<1K"
    elif total < 10000:
        size_category = "1K<n<10K"
    elif total < 100000:
        size_category = "10K<n<100K"
    elif total < 1000000:
        size_category = "100K<n<1M"
    else:
        size_category = "1M<n<10M"

    generation_date = datetime.now().strftime("%Y-%m-%d")

    card_content = f'''---
language:
- vi
license: cc-by-sa-4.0
task_categories:
- text-generation
- fill-mask
- text-classification
- feature-extraction
- sentence-similarity
tags:
- wikipedia
- vietnamese
- nlp
- underthesea
- wikidata
- pretraining
- language-modeling
pretty_name: UVW 2026 - Vietnamese Wikipedia Dataset
size_categories:
- {size_category}
source_datasets:
- original
dataset_info:
  features:
  - name: id
    dtype: string
  - name: title
    dtype: string
  - name: content
    dtype: string
  - name: num_chars
    dtype: int32
  - name: num_sentences
    dtype: int32
  - name: quality_score
    dtype: int32
  - name: wikidata_id
    dtype: string
  - name: main_category
    dtype: string
  splits:
  - name: train
    num_examples: {train_count if isinstance(train_count, int) else 0}
  - name: validation
    num_examples: {val_count if isinstance(val_count, int) else 0}
  - name: test
    num_examples: {test_count if isinstance(test_count, int) else 0}
configs:
- config_name: default
  data_files:
  - split: train
    path: train.parquet
  - split: validation
    path: validation.parquet
  - split: test
    path: test.parquet
---

# UVW 2026: Underthesea Vietnamese Wikipedia Dataset

<div align="center">

[![License: CC BY-SA 4.0](https://img.shields.io/badge/License-CC%20BY--SA%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-sa/4.0/)
[![Language: Vietnamese](https://img.shields.io/badge/Language-Vietnamese-blue.svg)](https://vi.wikipedia.org)
[![Wikidata Enriched](https://img.shields.io/badge/Wikidata-Enriched-green.svg)](https://www.wikidata.org)

</div>

## Dataset Description

**UVW 2026** (Underthesea Vietnamese Wikipedia) is a high-quality, cleaned dataset of Vietnamese Wikipedia articles enriched with Wikidata metadata. Designed for Vietnamese NLP research including language modeling, text generation, text classification, named entity recognition, and model pretraining.

### Key Features

- **Clean text**: Wikipedia markup, templates, references, and formatting removed
- **Wikidata integration**: Articles linked to Wikidata entities with semantic categories
- **Quality scoring**: Each article scored 1-10 based on content quality metrics
- **Unicode normalized**: NFC normalization applied for consistent text processing
- **Ready to use**: Pre-split into train/validation/test sets

### Dataset Summary

| Property | Value |
|----------|-------|
| **Language** | Vietnamese (vi) |
| **Source** | Vietnamese Wikipedia + Wikidata |
| **License** | CC BY-SA 4.0 |
| **Generated** | {generation_date} |
| **Total Articles** | {total:,} |
| **Wikidata Coverage** | {wikidata_pct:.1f}% |
| **Category Coverage** | {category_pct:.1f}% |
| **Unique Categories** | {unique_categories:,} |
| **Avg. Characters** | {avg_chars:,} |
| **Avg. Sentences** | {avg_sentences:,} |

## Quick Start

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("undertheseanlp/UVW-2026")

# Access splits
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

# View an example
print(train[0])
```

## Dataset Structure

### Data Splits

| Split | Examples | Description |
|-------|----------|-------------|
| `train` | {train_count:,} | Training set (80%) |
| `validation` | {val_count:,} | Validation set (10%) |
| `test` | {test_count:,} | Test set (10%) |

### Schema

```json
{{
  "id": "Việt_Nam",
  "title": "Việt Nam",
  "content": "Việt Nam, tên chính thức là Cộng hòa Xã hội chủ nghĩa Việt Nam...",
  "num_chars": 45000,
  "num_sentences": 500,
  "quality_score": 9,
  "wikidata_id": "Q881",
  "main_category": "quốc gia có chủ quyền"
}}
```

### Field Descriptions

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique article identifier (URL-safe title) |
| `title` | string | Human-readable article title |
| `content` | string | Cleaned article text content |
| `num_chars` | int32 | Character count of content |
| `num_sentences` | int32 | Estimated sentence count |
| `quality_score` | int32 | Quality score from 1 (lowest) to 10 (highest) |
| `wikidata_id` | string | Wikidata Q-identifier (e.g., "Q881" for Vietnam) |
| `main_category` | string | Primary category from Wikidata P31 (instance of) |

## Usage Examples

### Filter High-Quality Articles

```python
# Get articles with quality score >= 7
high_quality = dataset["train"].filter(lambda x: x["quality_score"] >= 7)
print(f"High-quality articles: {{len(high_quality):,}}")
```

### Filter by Category

```python
# Get articles about people
people = dataset["train"].filter(lambda x: x["main_category"] == "người")
print(f"Articles about people: {{len(people):,}}")

# Get articles about locations
locations = dataset["train"].filter(
    lambda x: "khu định cư" in (x["main_category"] or "")
)
```

### Filter by Wikidata

```python
# Get articles with Wikidata links
with_wikidata = dataset["train"].filter(lambda x: x["wikidata_id"] != "")

# Lookup specific entity
vietnam = dataset["train"].filter(lambda x: x["wikidata_id"] == "Q881")
```

### Use for Language Modeling

```python
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("vinai/phobert-base")

def tokenize(examples):
    return tokenizer(examples["content"], truncation=True, max_length=512)

tokenized = dataset["train"].map(tokenize, batched=True)
```

## Quality Score

Articles are scored 1-10 based on multiple factors:

| Component | Weight | Criteria |
|-----------|--------|----------|
| **Length** | 40% | Character count (200 - 100,000 optimal) |
| **Sentences** | 30% | Sentence count (3 - 1,000 optimal) |
| **Density** | 30% | Avg sentence length (80-150 chars optimal) |
| **Wikidata bonus** | +0.5 | Has wikidata_id |
| **Category bonus** | +0.5 | Has main_category |
| **Markup penalty** | -1 to -3 | Remaining Wikipedia markup |

### Quality Distribution

| Score | Count | Percentage |
|-------|------:|----------:|
{quality_table}

## Top Categories

| Category (Vietnamese) | Count | Percentage |
|----------------------|------:|----------:|
{top_categories_table}

## Data Processing

### Pipeline Steps

1. **Download**: Fetch Vietnamese Wikipedia XML dump from Wikimedia
2. **Extract**: Parse XML and extract article content
3. **Clean**: Remove Wikipedia markup (templates, refs, links, tables, categories)
4. **Normalize**: Apply Unicode NFC normalization
5. **Score**: Calculate quality metrics for each article
6. **Enrich**: Add Wikidata IDs and semantic categories via Wikidata API
7. **Filter**: Remove special pages, redirects, disambiguation, and short articles (<100 chars)
8. **Split**: Create train/validation/test splits (80/10/10) with seed=42

### Removed Content

- Wikipedia templates (`{{{{...}}}}`)
- References and citations (`<ref>...</ref>`)
- HTML tags and comments
- Category links (`[[Thể loại:...]]`)
- File/image links (`[[Tập tin:...]]`, `[[File:...]]`)
- Interwiki links
- Tables (`{{| ... |}}`)
- Infoboxes and navigation templates

### Reproduction

```bash
git clone https://github.com/undertheseanlp/UVW-2026
cd UVW-2026
uv sync --extra huggingface

# Run full pipeline
uv run python scripts/build_dataset.py

# Or run individual steps
uv run python scripts/download_wikipedia.py
uv run python scripts/extract_articles.py
uv run python scripts/wikipedia_quality_score.py
uv run python scripts/add_wikidata.py
uv run python scripts/create_splits.py
uv run python scripts/prepare_huggingface.py --push
```

## Citation

```bibtex
@dataset{{uvw2026,
  title = {{UVW 2026: Underthesea Vietnamese Wikipedia Dataset}},
  author = {{Underthesea NLP}},
  year = {{2026}},
  publisher = {{Hugging Face}},
  url = {{https://huggingface.co/datasets/undertheseanlp/UVW-2026}},
  note = {{Vietnamese Wikipedia articles enriched with Wikidata metadata}}
}}
```

## Related Resources

- [Underthesea](https://github.com/undertheseanlp/underthesea) - Vietnamese NLP Toolkit
- [PhoBERT](https://github.com/VinAIResearch/PhoBERT) - Pre-trained language models for Vietnamese
- [Vietnamese Wikipedia](https://vi.wikipedia.org)
- [Wikidata](https://www.wikidata.org)

## License

This dataset is released under the [Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0)](https://creativecommons.org/licenses/by-sa/4.0/), consistent with the Wikipedia content license.

---

<div align="center">
Made with ❤️ by <a href="https://github.com/undertheseanlp">Underthesea NLP</a>
</div>
'''

    HF_DIR.mkdir(parents=True, exist_ok=True)

    # Save to main HF directory
    readme_path = HF_DIR / "README.md"
    with open(readme_path, "w", encoding="utf-8") as f:
        f.write(card_content)
    print(f"  Dataset card saved to: {readme_path}")

    # Also save to parquet directory for Hub upload
    parquet_readme = HF_DIR / "uvw_2026_parquet" / "README.md"
    if parquet_readme.parent.exists():
        with open(parquet_readme, "w", encoding="utf-8") as f:
            f.write(card_content)
        print(f"  Dataset card saved to: {parquet_readme}")


def push_to_hub(repo_id: str, private: bool = False):
    """Push dataset to HuggingFace Hub."""
    if not HF_HUB_AVAILABLE:
        print("Error: huggingface-hub library not installed")
        print("Install with: uv sync --extra huggingface")
        return False

    parquet_dir = HF_DIR / "uvw_2026_parquet"
    if not parquet_dir.exists():
        print(f"Error: Parquet directory not found: {parquet_dir}")
        print("Please run without --push first to generate the dataset.")
        return False

    print(f"\nPushing to HuggingFace Hub: {repo_id}")

    try:
        api = HfApi()

        # Check if user is logged in
        try:
            user_info = api.whoami()
            print(f"  Authenticated as: {user_info['name']}")
        except Exception:
            print("  Not logged in. Please run: huggingface-cli login")
            return False

        # Create or update the repository
        print(f"  Creating/updating repository...")
        api.create_repo(
            repo_id=repo_id,
            repo_type="dataset",
            private=private,
            exist_ok=True,
        )

        # Upload all files in parquet directory
        print(f"  Uploading files from {parquet_dir}...")
        api.upload_folder(
            folder_path=str(parquet_dir),
            repo_id=repo_id,
            repo_type="dataset",
            commit_message="Update UVW 2026 dataset",
        )

        print(f"\n  Successfully pushed to: https://huggingface.co/datasets/{repo_id}")
        return True

    except Exception as e:
        print(f"Error pushing to Hub: {e}")
        return False


def parse_args():
    """Parse command line arguments."""
    parser = argparse.ArgumentParser(
        description="Prepare UVW 2026 dataset for HuggingFace Hub",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
  # Generate dataset files only
  python scripts/prepare_huggingface.py

  # Generate and push to Hub
  python scripts/prepare_huggingface.py --push

  # Push to a custom repository
  python scripts/prepare_huggingface.py --push --repo-id myorg/my-dataset

  # Push as private dataset
  python scripts/prepare_huggingface.py --push --private

  # Only push (skip regeneration)
  python scripts/prepare_huggingface.py --push --skip-generate
        """,
    )

    parser.add_argument(
        "--push",
        action="store_true",
        help="Push dataset to HuggingFace Hub after generation",
    )
    parser.add_argument(
        "--repo-id",
        type=str,
        default="undertheseanlp/UVW-2026",
        help="HuggingFace repository ID (default: undertheseanlp/UVW-2026)",
    )
    parser.add_argument(
        "--private",
        action="store_true",
        help="Create private repository on Hub",
    )
    parser.add_argument(
        "--skip-generate",
        action="store_true",
        help="Skip dataset generation, only push existing files",
    )
    parser.add_argument(
        "--stats",
        action="store_true",
        help="Calculate detailed statistics for dataset card",
    )

    return parser.parse_args()


def main():
    """Prepare dataset for HuggingFace Hub."""
    args = parse_args()

    print("=" * 60)
    print("UVW 2026 - HuggingFace Dataset Preparation")
    print("=" * 60)

    split_counts = None
    stats = None

    if not args.skip_generate:
        # Check dependencies
        if not HF_AVAILABLE:
            print("\nError: datasets library not installed")
            print("Install with: uv sync --extra huggingface")
            return

        # Check if splits exist
        if not SPLITS_DIR.exists():
            print(f"\nError: Splits directory not found: {SPLITS_DIR}")
            print("Please run create_splits.py first.")
            return

        # Calculate statistics if requested
        if args.stats:
            print("\nCalculating dataset statistics...")
            stats = load_dataset_statistics()

        # Create dataset
        print("\nCreating HuggingFace dataset...")
        result = create_hf_dataset()

        if result:
            _, split_counts = result

        # Create dataset card
        print("\nCreating dataset card...")
        create_dataset_card(split_counts, stats)

        print("\nDataset preparation complete!")

    # Push to Hub if requested
    if args.push:
        success = push_to_hub(args.repo_id, private=args.private)
        if not success:
            return

    # Show next steps if not pushing
    if not args.push:
        print("\n" + "-" * 60)
        print("Next steps:")
        print("-" * 60)
        print("\nTo upload to HuggingFace Hub, run:")
        print(f"  python scripts/prepare_huggingface.py --push")
        print("\nOr upload manually:")
        print("  1. huggingface-cli login")
        print("  2. huggingface-cli upload undertheseanlp/UVW-2026 data/huggingface/uvw_2026_parquet")

    print("\nDone!")


if __name__ == "__main__":
    main()