LLM-texts-gutenberg

Summary

LLM-texts-gutenberg is a large-scale collection of public domain texts sourced from Project Gutenberg. The dataset consists of full-text ebooks that have been cleaned, formatted, and prepared for large language model (LLM) training and other natural language processing (NLP) tasks.

With 2,622,640,269 total tokens, this dataset provides a rich corpus of classic literature, historical documents, and cultural works spanning a wide range of genres, authors, and time periods. The text has been cleaned to remove common noise artifacts such as bracketed numbers, links, stray letters, and similar formatting/OCR noise.

The dataset is provided in two formats:

1. JSONL – Each line contains a full ebook with metadata and the cleaned plain text.
2. Pre-tokenized Binary – For efficient memory-mapped loading, the dataset includes multiple binary files named llm_dataset_{TOKENIZER_NAME}_tokenizer.bin. These contain the tokenized representations of the entire corpus, tokenized with popular tokenizers (e.g., gpt2, llama, bert, etc.). This format is ideal for large-scale training without repetitive tokenization overhead.

Supported Tasks and Leaderboards

• Language Modeling – The dataset is well-suited for pre-training or fine-tuning LLMs on diverse, high-quality English text.
• Text Generation – Can be used to train models for creative writing, story generation, or stylistic imitation.
• Document Classification – The variety of genres and authors enables classification tasks (e.g., authorship attribution, genre prediction).
• Question Answering & Reading Comprehension – With appropriate preprocessing, the dataset can support extractive QA and comprehension tasks.

Languages

The dataset primarily contains English texts, as Project Gutenberg's collection is predominantly in English. However, a small number of works in other languages may also be present.

---

Dataset Structure

Data Instances

Each record in the JSONL corresponds to a single ebook from Project Gutenberg and includes the following fields:

{
  "url": "https://www.gutenberg.org/cache/epub/32453/pg32453-images.html",
  "license": "Public Domain (Project Gutenberg)",
  "license_verified": true,
  "fetched_at": "2026-09-06T06:11:57",
  "content_hash": "20fe50c25af6a06c310294519d785d41bc4bd4643bc538ca55e1bf0ceb0c4eee",
  "text": "Author of\"The Tucker Twins Series,\" \"The CarterGirls Series,\" etc.\nMolly Brown's Sophomore (...TRUNCATED)"
}

Data Fields

Field	Type	Description
url	string	The source URL of the ebook on Project Gutenberg.
license	string	The license under which the ebook is released (typically Public Domain (Project Gutenberg)).
license_verified	bool	Whether the license has been programmatically verified.
fetched_at	timestamp	The date and time when the ebook was downloaded from Project Gutenberg.
content_hash	string	A SHA-256 hash of the ebook content for deduplication and integrity checks.
text	string	The cleaned full text of the ebook, including front matter, main content, and back matter where retained after noise removal.

Tokenized Binary Format

The dataset includes one or more binary files named llm_dataset_{TOKENIZER_NAME}_tokenizer.bin. These files contain the entire corpus encoded as a flat array of token IDs, stored in a memory-mappable binary format (e.g., using numpy or torch tensors). This allows users to load the dataset without re-tokenizing, saving significant RAM and CPU time during training.

The TOKENIZER_NAME in the filename indicates which tokenizer was used (e.g., gpt2, llama, bert, roberta, etc.). Each file corresponds to a specific tokenizer and includes:

• The total number of tokens for that tokenizer.
• The token IDs in a contiguous binary array, ready for streaming or random access.

Dataset Statistics

Metric	Value
Total Tokens	2,622,640,269 tokens
Number of Ebooks	(not specified, but likely tens of thousands)
Raw Format	JSONL (line-delimited JSON, cleaned text)
Tokenized Format	Binary files (.bin) for each supported tokenizer
Available Tokenizers	gpt2, llama, bert (and possibly others – see filenames in the repository)
Cleaning	Removed common noise artifacts such as bracketed numbers, links, stray letters, and similar formatting/OCR noise

Token count was provided by the dataset creator after cleaning. The exact tokenization methodology used for counting is not specified, but it is assumed to be based on a standard subword tokenizer (e.g., GPT-2).

---

Dataset Creation

Curation Rationale

The dataset was created to provide a large, freely available corpus of high-quality public domain texts for training large language models. By aggregating texts from Project Gutenberg, the dataset ensures that all content is legally usable for both research and commercial applications.

Source Data

• Project Gutenberg (gutenberg.org) – A digital library of over 70,000 free ebooks, primarily works whose U.S. copyright has expired.
• The dataset likely includes a broad selection of ebooks, covering fiction, non-fiction, poetry, drama, and historical documents.

Data Processing

The texts were fetched from Project Gutenberg's HTML cache, parsed, and stored as plain text. Each ebook is stored as a separate record with its metadata. The extracted text was cleaned to remove common noise artifacts such as bracketed numbers, links, stray letters, and similar formatting/OCR noise. HTML markup was stripped to retain only the readable text.

For the binary tokenized versions, the cleaned text was tokenized using various standard tokenizers, and the resulting token ID sequences were concatenated into a single large binary array. This preprocessing step is performed offline to enable faster loading during model training.

---

Considerations for Using the Data

Social Impact of the Dataset

• Cultural Heritage: The dataset preserves and makes accessible a vast collection of human literary and historical works.
• Bias and Representation: The corpus is heavily weighted toward Western, English-language authors, with an overrepresentation of works from the 19th and early 20th centuries. This may introduce historical and cultural biases into any model trained on it.
• Outdated Content: Some texts may contain language, viewpoints, or depictions that are offensive or outdated by modern standards. Users should be aware of this when training models.

Discussion of Biases

• Genre Imbalance: The dataset likely includes more fiction and poetry than technical or non-fiction works.
• Author Demographics: The majority of authors are Western, male, and writing in English. This reflects the historical composition of Project Gutenberg's collection.

Other Known Limitations

• Quality Variances: Some ebooks may contain residual OCR errors, formatting inconsistencies, or missing sections.
• Duplication: The dataset may include multiple editions or versions of the same work.
• Noisy Metadata: While most entries have a valid license, some records may have incomplete or inconsistent metadata.

---

Additional Information

Dataset Curators

The dataset was curated and uploaded by sirunchained on Hugging Face.

Licensing Information

The texts in this dataset were sourced from Project Gutenberg. The underlying works are generally identified by Project Gutenberg as being in the public domain in the United States. Public-domain status may vary by jurisdiction.

The dataset compilation, including its organization, metadata, and pre-tokenized files, is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Users of this dataset are responsible for determining whether the use of individual works is permitted under the laws applicable to them.

For the source and rights information of an individual work, please refer to its corresponding Project Gutenberg record.

Citation

If you use this dataset in your research, please consider citing it as:

@dataset{sirunchained_llm_texts_gutenberg,
  author       = {sirunchained},
  title        = {LLM-texts-gutenberg},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/sirunchained/LLM-texts-gutenberg},
  note         = {Total tokens: 2,622,640,269. Dual-licensed: Public Domain texts, CC BY 4.0 for compilation.}
}

Contributions

Contributions to improve the dataset (e.g., cleaning, deduplication, or adding metadata) are welcome. Please open a discussion or pull request on the Hugging Face repository.