Datasets:

NNEngine
/

Gutenberg-Clean-40M

license: mit
task_categories:
  - text-classification
language:
  - en
tags:
  - text-generation
  - causal
  - training
  - transformers
  - pytorch
  - jsonl
  - segmentation
  - validation
size_categories:
  - 10M<n<100M

📚 TinyWay-Gutenberg-Clean-40M

A large-scale, high-quality English text dataset derived from Project Gutenberg, cleaned, normalized, deduplicated, and segmented into fixed-length samples for efficient language model pretraining.

This dataset is designed to support training small and medium language models such as TinyWay, tokenizer training, embedding models, and large-scale NLP experimentation.

Dataset Overview

Name: TinyWay-Gutenberg-Clean-40M
Samples: ~40,000,000
Language: English
Format: JSONL (optionally gzip-compressed)
Source: Project Gutenberg (public domain books)
License: Public Domain
Intended Use: Language model pretraining, tokenizer training, representation learning

Each line in the dataset contains a clean text segment between 30 and 60 words.

Data Format

Each record is stored as a JSON object:

{
  "id": "twg_000000000123",
  "text": "Cleaned text segment of natural English language between thirty and sixty words.",
  "word_count": 42,
  "source": "gutenberg"
}

Fields

Field	Description
`id`	Unique sample identifier
`text`	Clean English text segment
`word_count`	Number of words in the sample
`source`	Data source identifier

Data Processing Pipeline

The dataset was generated using a fully streaming pipeline to ensure scalability and low memory usage.

Steps

Streaming Input
- Data loaded from a Project Gutenberg mirror using Hugging Face streaming APIs.
Text Cleaning
- Removed Gutenberg headers and footers
- Removed chapter titles and page numbers
- Normalized whitespace and line breaks
- Removed non-ASCII and control characters
- Removed URLs and artifacts
Segmentation
- Text split into fixed segments of 30–60 words.
Validation
- Enforced word count constraints
- Filtered short or malformed segments
Deduplication
- Exact hash-based deduplication applied during generation.
Output
- Stored as JSONL files (optionally gzip-compressed).
- Sharded for easier distribution and loading.

How to Load the Dataset

Using Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset(
    "NNEngine/TinyWay-Gutenberg-Clean-40M",
    split="train",
    streaming=True
)

for sample in dataset.take(3):
    print(sample)

Reading JSONL Manually

import json

with open("data/train-00000.jsonl", "r", encoding="utf-8") as f:
    for _ in range(3):
        print(json.loads(next(f)))

If files are compressed:

import gzip
import json

with gzip.open("train-00000.jsonl.gz", "rt", encoding="utf-8") as f:
    for _ in range(3):
        print(json.loads(next(f)))

Dataset Characteristics

Approximate properties:

Average words per sample: ~45
Vocabulary: Large natural English vocabulary
Style: Literary and narrative English
Domain: Fiction, non-fiction, historical texts

Limitations

Content is primarily literary and historical in nature.
No conversational, chat, or code data.
Some archaic vocabulary and sentence structure may appear.
Deduplication is hash-based (near-duplicates may remain).

For conversational or modern web text, additional datasets should be mixed.

License

All source texts originate from Project Gutenberg and are in the public domain. This processed dataset is released for unrestricted research and commercial use.

Citation

If you use this dataset in research or publications, please cite:

TinyWay-Gutenberg-Clean-40M
NNEngine, 2026

🧠 Maintainer

Created and maintained by Shivam Sharma