longcrawl64.py

import os
import zarr
import numpy as np
import datasets

logger = datasets.logging.get_logger(__name__)

_CITATION = """\
@misc{buckman2024,
    author = {Buckman, Jacob},
    publisher = {Manifest AI},
    title = {LongCrawl64: {A} {Long-Context} {Natural-Language} {Dataset}},
    date = {2024-08-14},
    langid = {en}
}
"""

_DESCRIPTION = """\
LongCrawl64 is a dataset for research on architectures and algorithms for long-context modeling.
It consists of 6,661,465 pre-tokenized documents, each of which is 65,536 tokens long, for a total
token count of 435 billion. The dataset is preprocessed with truncation to exactly 64 KiT,
shuffling along document dimension, and rolling each document randomly along sequence dimension.
"""


class LongCrawl64Config(datasets.BuilderConfig):
    """BuilderConfig for LongCrawl64."""

    def __init__(self, context_size=65536, **kwargs):
        """BuilderConfig for LongCrawl64.

        Args:
            context_size: The size of context window to use (default is full 64KiT)
            **kwargs: keyword arguments forwarded to super.
        """
        super().__init__(version=datasets.Version("1.0.0"), **kwargs)
        self.context_size = context_size


class LongCrawl64(datasets.GeneratorBasedBuilder):
    """LongCrawl64 dataset."""

    BUILDER_CONFIGS = [
        LongCrawl64Config(
            name="default",
            description="Default configuration with full 64KiT context",
        ),
        LongCrawl64Config(
            name="16k",
            description="16K context window configuration",
            context_size=16384,
        ),
    ]

    DEFAULT_CONFIG_NAME = "default"

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "tokens": datasets.Sequence(
                        datasets.Value("int32"), length=self.config.context_size
                    ),
                    "input_ids": datasets.Sequence(
                        datasets.Value("int32"), length=self.config.context_size
                    ),
                }
            ),
            supervised_keys=None,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        # Download the data files from HuggingFace Hub
        data_files = {"train": "data/train.zarr", "validation": "data/heldout.zarr"}

        downloaded_files = dl_manager.download_and_extract(data_files)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "zarr_path": downloaded_files["train"],
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "zarr_path": downloaded_files["validation"],
                },
            ),
        ]

    def _generate_examples(self, zarr_path):
        """Yields examples.

        Reads data from the zarr store in chunks and yields examples
        according to the specified context size.
        """
        logger.info(f"Loading zarr array from {zarr_path}")
        z = zarr.open(zarr_path, mode="r")

        # Get the first array in the zarr store (0.0)
        data = z["0.0"]

        # Calculate number of complete sequences we can get from each document
        seqs_per_doc = data.shape[1] // self.config.context_size

        for doc_idx in range(data.shape[0]):
            # Read the document data
            doc_data = data[doc_idx]

            for seq_idx in range(seqs_per_doc):
                # Extract sequence
                start = seq_idx * self.config.context_size
                end = start + self.config.context_size
                sequence = doc_data[start:end]

                # Create input_ids by shifting sequence
                input_ids = np.roll(sequence, 1)
                input_ids[0] = 50256  # EOT token as per the paper

                yield f"{doc_idx}-{seq_idx}", {
                    "tokens": sequence.tolist(),
                    "input_ids": input_ids.tolist(),
                }