lm1b_builder.py

import os


class LM1B:
    def __init__(self, root_path: str = "dataset/lm1b"):
        self._root_path: str = root_path
        self._tf_path = os.path.join(root_path, "tf")
        self._hf_path = os.path.join(root_path, "hf")

    def download(self):
        """Download LM1B dataset from TensorFlow Datasets.

        Downloads the dataset and saves it under ${self._root_path}/tf
        Reference: https://www.tensorflow.org/datasets/catalog/lm1b
        """
        import tensorflow_datasets as tfds

        # Create the tf directory if it doesn't exist
        os.makedirs(self._tf_path, exist_ok=True)

        # Download LM1B dataset using TFDS
        # This will download and prepare the dataset
        tfds.load(
            "lm1b",
            data_dir=self._tf_path,
            download=True,
        )
        print(f"LM1B dataset downloaded to {self._tf_path}")

    def convert_to_hf(self):
        """Load downloaded LM1B with TFDS and convert it to HuggingFace Dataset format.

        Saves the HuggingFace version under ${self._root_path}/hf
        """
        import tensorflow_datasets as tfds
        from datasets import Dataset, DatasetDict

        # Create the hf directory if it doesn't exist
        os.makedirs(self._hf_path, exist_ok=True)

        # Load the dataset from the local tf directory
        ds_builder = tfds.builder("lm1b", data_dir=self._tf_path)

        # Get available splits
        splits = ds_builder.info.splits

        hf_datasets = {}

        for split_name in splits:
            print(f"Converting {split_name} split...")

            # Load the TF dataset
            tf_dataset = tfds.load(
                "lm1b",
                split=split_name,
                data_dir=self._tf_path,
                download=False,
            )

            # Convert to list of dictionaries
            texts = []
            for example in tf_dataset:
                # LM1B has a 'text' field
                text = example["text"].numpy().decode("utf-8")
                texts.append({"text": text})

            # Create HuggingFace Dataset
            hf_datasets[split_name] = Dataset.from_list(texts)
            print(f"  Converted {len(texts)} examples")

        # Create DatasetDict and save
        dataset_dict = DatasetDict(hf_datasets)
        dataset_dict.save_to_disk(self._hf_path)
        print(f"HuggingFace dataset saved to {self._hf_path}")

    def add_meta_data(self):
        """Add detailed metadata to the converted dataset.

        Specifies entry field, split information, and dataset card.
        """
        from datasets import load_from_disk
        import json

        # Load the HF dataset
        dataset = load_from_disk(self._hf_path)

        # Create metadata
        metadata = {
            "name": "lm1b",
            "description": "One Billion Word Benchmark for Language Modeling. "
                           "A benchmark corpus for measuring progress in statistical language modeling.",
            "citation": """@inproceedings{chelba2013one,
  title={One billion word benchmark for measuring progress in statistical language modeling},
  author={Chelba, Ciprian and Mikolov, Tomas and Schuster, Mike and Ge, Qi and Brants, Thorsten and Koehn, Phillipp and Robinson, Tony},
  booktitle={Interspeech},
  year={2014}
}""",
            "homepage": "https://www.statmt.org/lm-benchmark/",
            "license": "Apache-2.0",
            "features": {
                "text": "string - The text content of each example"
            },
            "splits": {k: {"num_examples": len(v)} for k, v in dataset.items()},
            "task_categories": ["text-generation", "fill-mask"],
            "language": ["en"],
            "size_category": "1B<n<10B",
        }

        # Save metadata
        metadata_path = os.path.join(self._hf_path, "metadata.json")
        with open(metadata_path, "w") as f:
            json.dump(metadata, f, indent=2)

        # Create a README dataset card
        readme_content = f"""---
license: apache-2.0
task_categories:
  - text-generation
  - fill-mask
language:
  - en
size_categories:
  - 1B<n<10B
---

# LM1B - One Billion Word Benchmark

## Dataset Description

The One Billion Word Benchmark is a large language modeling dataset.
It contains approximately one billion words of training data derived from news articles.

## How was this dataset built?

We download the full LM1B dataset from TensorFlow Datasets (TFDS) and convert it to HuggingFace format automatically. The full script is in `lm1b.py`. The required environment is:
- tensorflow==2.20.0
- tensorflow-datasets==4.9.9
- huggingface_hub==1.3.3
- datasets==4.4.1

```bash
pip install tensorflow==2.20.0 tensorflow-datasets==4.9.9 huggingface_hub==1.3.3 datasets==4.4.1
python lm1b_builder.py --action all
```

## Dataset Structure

### Data Fields

- `text`: A string containing the text content

### Data Splits

| Split | Examples |
|-------|----------|
"""
        for split_name, split_data in dataset.items():
            readme_content += f"| {split_name} | {len(split_data):,} |\n"

        readme_content += """
## Citation

```bibtex
@inproceedings{chelba2013one,
  title={One billion word benchmark for measuring progress in statistical language modeling},
  author={Chelba, Ciprian and Mikolov, Tomas and Schuster, Mike and Ge, Qi and Brants, Thorsten and Koehn, Phillipp and Robinson, Tony},
  booktitle={Interspeech},
  year={2014}
}
```

## License

Apache 2.0
"""
        readme_path = os.path.join(self._hf_path, "README.md")
        with open(readme_path, "w") as f:
            f.write(readme_content)

        print(f"Metadata saved to {metadata_path}")
        print(f"README saved to {readme_path}")

    def upload_to_hf(self, user: str, name: str):
        """Upload the converted dataset to HuggingFace Hub.

        Args:
            user: HuggingFace username or organization
            name: Repository name for the dataset
        """
        from datasets import load_from_disk
        from huggingface_hub import HfApi

        # Load the dataset
        dataset = load_from_disk(self._hf_path)

        # Create the repo name
        repo_id = f"{user}/{name}"

        # Push to hub
        dataset.push_to_hub(
            repo_id,
            private=False,
        )

        # Upload additional files (README.md, metadata.json)
        api = HfApi()

        readme_path = os.path.join(self._hf_path, "README.md")
        if os.path.exists(readme_path):
            api.upload_file(
                path_or_fileobj=readme_path,
                path_in_repo="README.md",
                repo_id=repo_id,
                repo_type="dataset",
            )

        metadata_path = os.path.join(self._hf_path, "metadata.json")
        if os.path.exists(metadata_path):
            api.upload_file(
                path_or_fileobj=metadata_path,
                path_in_repo="metadata.json",
                repo_id=repo_id,
                repo_type="dataset",
            )

        # Upload this script itself for reproducibility
        script_path = os.path.abspath(__file__)
        if os.path.exists(script_path):
            api.upload_file(
                path_or_fileobj=script_path,
                path_in_repo="lm1b_builder.py",
                repo_id=repo_id,
                repo_type="dataset",
            )

        print(f"Dataset uploaded to https://huggingface.co/datasets/{repo_id}")

    def test(self):
        """Test if the converted HuggingFace dataset can be loaded correctly."""
        from datasets import load_from_disk

        print("Testing local HuggingFace dataset...")

        # Load from disk
        dataset = load_from_disk(self._hf_path)

        print(f"Dataset loaded successfully!")
        print(f"Available splits: {list(dataset.keys())}")

        for split_name, split_data in dataset.items():
            print(f"\n{split_name} split:")
            print(f"  Number of examples: {len(split_data):,}")
            print(f"  Features: {split_data.features}")

            # Show a sample
            if len(split_data) > 0:
                sample = split_data[0]
                text_preview = sample["text"][:100] + "..." if len(sample["text"]) > 100 else sample["text"]
                print(f"  Sample text: {text_preview}")

        print("\nAll tests passed!")
        return True
    
    def download_convert_metadata_test(self):
        """Unified method that runs the full pipeline: download, convert, add metadata, and test."""
        print("=== Step 1: Download ===")
        self.download()
        print("\n=== Step 2: Convert to HuggingFace ===")
        self.convert_to_hf()
        print("\n=== Step 3: Add Metadata ===")
        self.add_meta_data()
        print("\n=== Step 4: Test ===")
        self.test()


if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(description="LM1B Dataset Preparation")
    parser.add_argument("--root_path", type=str, default="dataset/lm1b", help="Root path for the dataset")
    parser.add_argument("--action", type=str, required=True,
                        choices=["download", "convert", "metadata", "upload", "test", "all"],
                        help="Action to perform")
    parser.add_argument("--user", type=str, help="HuggingFace username (for upload)")
    parser.add_argument("--name", type=str, help="Dataset name on HuggingFace (for upload)")

    args = parser.parse_args()

    lm1b = LM1B(root_path=args.root_path)

    if args.action == "download":
        lm1b.download()
    elif args.action == "convert":
        lm1b.convert_to_hf()
    elif args.action == "metadata":
        lm1b.add_meta_data()
    elif args.action == "upload":
        if not args.user or not args.name:
            raise ValueError("--user and --name are required for upload")
        lm1b.upload_to_hf(args.user, args.name)
    elif args.action == "test":
        lm1b.test()
    elif args.action == "all":
        lm1b.download_convert_metadata_test()
        print("\nTo upload, run with --action upload --user <user> --name <name>")