meettilavat/Gutenberg_1899

Gutenberg Public Domain Text Corpus

Cleaned snapshot of 57,117 public-domain works from Project Gutenberg obtained through the Gutendex API. Each document is the body text of a single title (authors deceased ≤ 1899), stored as UTF‑8 strings in 33 Zstd-compressed Parquet shards. Boilerplate headers/footers are stripped, whitespace is lightly normalized, and texts shorter than 1,000 characters or with an ASCII ratio below 50 % are removed.

• Documents: 57,117
• Characters: 22,122,271,753 (≈5.5 B tokens using a 4-char heuristic)
• Average length: 387,315 characters
• Median length: 308,079 characters
• 95th percentile length: 1,020,612 characters
• Languages: Primarily English with additional Western European languages passing the ASCII threshold.

Files

• shard_00000.parquet … shard_00032.parquet: Book texts in a single text column, Zstd level 1, row groups of 1,024 rows. Each shard targets ~5 × 10⁸ characters.
• processing_report.txt (optional): Collection statistics and throughput metrics generated during the download.
• checkpoint_processed_books.txt (optional): List of Gutenberg IDs already processed for resumable runs.

Dataset Structure

Data Instances

Each row consists of a full book after cleaning:

{
  "text": "AMERICAN NOTES FOR GENERAL CIRCULATION ... (full book text)"
}

Data Fields

• text (string): UTF-8 encoded body content without Gutenberg boilerplate. Maximum observed length 27,942,687 characters.

Data Splits

No predefined splits. All documents belong to a single training split; define custom splits as needed.

Usage

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={"train": "shard_*.parquet"},
    split="train"
)

print(len(dataset))          # 57117
print(dataset[0]["text"][:500])

PyArrow

import pyarrow.parquet as pq

table = pq.read_table("shard_00000.parquet")
texts = table.column("text").to_pylist()

Fine-tuning Example (Transformers)

from datasets import load_dataset
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-125M")
dataset = load_dataset("parquet", data_files={"train": "shard_*.parquet"}, split="train")

def preprocess(batch):
    return tokenizer(batch["text"], truncation=True, padding="max_length", max_length=1024)

tokenized = dataset.map(preprocess, batched=True, remove_columns=["text"])

Reproducing the Dataset

To recreate this corpus, iterate over https://gutendex.com/books/?author_year_end=1899, download the UTF-8 plain text for each title, strip the standard Project Gutenberg boilerplate markers, collapse repeated whitespace (four or more consecutive newlines → triple newlines), filter out texts shorter than 1,000 characters or with fewer than 50 % ASCII characters, and batch the remaining documents into Parquet shards targeting ~500 M characters each using Zstandard compression. A resumable, high-concurrency Python pipeline with requests, pyarrow, tqdm, and psutil was used for the original collection, but any equivalent tooling can be substituted.

Considerations

Social Impact

Provides an accessible corpus of long-form prose for open research and language-model experimentation, especially for extended context training.

Biases and Limitations

• Dominated by Western canon authors whose works entered the public domain by 1899; social views and representation are limited and dated.
• Minimal language filtering means some non-English texts may be truncated if diacritics trigger the ASCII ratio filter.
• No metadata (title, author, year) accompanies the texts; merge Gutendex metadata separately if required.
• Duplicate editions or translations may appear because no deduplication beyond Gutendex cataloging was performed.

License

Works are believed to be in the public domain in the United States. When redistributing, acknowledge Project Gutenberg and review their terms of use. Ensure compliance with copyright law in your jurisdiction before using the dataset commercially.

Citation

• Project Gutenberg. Project Gutenberg [Data set]. https://www.gutenberg.org/
• Albert, K. (2020). Gutendex [API]. https://gutendex.com/

Maintainers

Dataset curated by the repository maintainer. For questions or issues, please open a ticket where the dataset is hosted or reach out via the provided contact channel.

Loading from Remote Hosting

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={"train": "https://path.to/shards/shard_*.parquet"}
)

Replace https://path.to/shards/ with the hosting URL of the Parquet shards.