haykgrigorian/english-historical-corpus-1800-1875

TimeCapsuleLLM World English 1800-1875

This dataset is the sharded pretraining corpus prepared for TimeCapsuleLLM-World-English-1800-1875-v1. It is a large English-language historical text collection centered on the years 1800-1875, with a mix of books, pamphlets, reports, newspapers, prose extracts, and other running text suitable for language-model pretraining.

The corpus was assembled from several cleaned source groups and then normalized into compressed JSONL shards for reliable transfer and downstream training.

What Is In This Release

• 164 compressed shard files in shards/*.jsonl.gz
• 6,759,267 total records
• 72,701,787,518 bytes of compressed shard output, about 72.7 GB
• 163,996,953,789 source-text bytes represented inside the shards, about 164.0 GB
• Estimated total size with the approved tokenizer: ~40.6 billion tokens

This shard set is the upload-ready representation of the cleaned corpus. The original copied package also contains manifests, inventories, tokenizer files, and validation outputs, but this dataset release focuses on the actual training text.

Date Range

• Primary target range: 1800-1875
• Chronicling America subset: 1800-1859

The project aimed to keep the corpus within the 1800-1875 window. Structured-source audits reported zero approved records outside the intended range for the staged sources included here.

Citation

If you use this dataset, please cite it as follows:

@misc{TimeCapsuleLLM World English 1800-1875,
  author = {Hayk Grigorian, Hamed Yaghoobian, Saber Soleymani},
  title = {World English (1800-1875)},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/haykgrigorian/TimeCapsuleLLM-World-English-1800-1875}}
}

Sources

1. London v2 cleaned text corpus

• Category: London-focused books and texts
• Records: 136,344
• Source bytes: 96,259,791,394
• Estimated tokens: 23,984,617,870
• Coverage: 1800-1875

This is the largest component of the corpus and provides the main London-centered historical English base for the model.

2. Multi-city cleaned corpora

• Category: Books and texts from 12 city folders
• Records: 109,490
• Source bytes: 55,186,381,124
• Estimated tokens: 13,146,827,944
• Coverage: 1800-1875

Cities included:

• Baltimore
• Belfast
• Birmingham
• Boston
• Chicago
• Dublin
• Edinburgh
• Liverpool
• Manchester
• New York
• Philadelphia
• Washington DC

3. Living with Machines strict newspaper subset

• Category: British newspapers
• Records: 535,882
• Source bytes: 2,857,330,295
• Estimated tokens: 793,662,715
• Coverage: 1800-1875

This subset was filtered strictly by date, OCR quality, and article length before being included.

4. Chronicling America language-filtered prose

• Category: US regional newspaper prose blocks
• Records: 903,599
• Source bytes: 1,255,276,477
• Estimated tokens: 357,171,910
• Coverage: 1800-1859

This source began as noisy OCR newspaper issues. It was converted into prose blocks and then filtered for advertisements, legal notices, merchant copy, and non-English content. Twelve original shard IDs produced no surviving prose and are represented by zero-byte shard placeholders in the source package, though the data included here contains only accepted records.

5. Internet Archive postfiltered prose

• Category: English prose blocks extracted from 1800-1875 Internet Archive texts
• Records: 5,073,952
• Source bytes: 8,438,174,499
• Estimated tokens: 2,310,063,013
• Coverage: 1800-1875

This source went through date, language, media-type, and deduplication filtering before paragraph-level prose extraction and targeted boilerplate stripping.

Living with Machines

• Strict date filtering
• OCR-quality filtering
• Minimum content-length filtering

Chronicling America

• Restricted issue selection
• Prose-block extraction from OCR issue text
• Advertisement, legal, merchant, and masthead filtering
• Language detection pass to remove non-English blocks

Internet Archive

• Date filtering to the 1800-1875 window
• English-language filtering
• Media-type filtering
• Front-matter and boilerplate stripping
• Newspaper and multi-column bleed rejection
• Deduplication against already collected corpus material
• Paragraph-level prose extraction and postfiltering

Shard Format

Each shard is a .jsonl.gz file. Each line is one JSON record with this general structure:

{
  "id": "source-specific-record-id",
  "text": "historical text content...",
  "metadata": {
    "source_name": "london_v2_cleaned_txt",
    "package_relative_source_path": "sources/...",
    "...": "additional source-specific fields"
  }
}

The exact metadata fields vary by source, but every record was normalized into the same top-level schema:

• id
• text
• metadata

Token Estimate

The corpus estimate is based on the approved v3 tokenizer, a HuggingFace Tokenizers ByteLevel BPE tokenizer trained from scratch for this broader corpus mix.

• Tokenizer vocab size: 32,000
• Special tokens: <s>, </s>, <unk>, <pad>, <|year|>, <|title|>
• Estimated total tokens in this corpus: 40.592B

The 72.7 GB dataset size on Hugging Face is the compressed shard size, not the raw-text size and not the token count.

Validation And Quality Notes

The build pipeline included:

• date-range audits
• boilerplate audits
• file inventory checks
• random readability spot checks

Project-level validation reported:

• 160/160 readability spot checks passed during the final package audit
• zero approved structured records outside the expected date range in the staged sources

This is still a historical OCR-heavy corpus, so some residual noise remains, especially in newspaper-derived material. The goal was not perfection but a large, usable historical English pretraining corpus with major non-prose debris removed.

Intended Use

This dataset is intended primarily for:

• historical language-model pretraining
• tokenizer analysis and training
• historical-domain continued pretraining
• corpus-level research on nineteenth-century English

It is not a benchmark dataset and it was not designed as a balanced annotation resource.

Limitations

• OCR noise is reduced, not eliminated
• Some sources are more book-heavy, others more newspaper-heavy
• Regional and institutional coverage is uneven
• Chronicling America is truncated at 1859 rather than the full 1875 range
• Historical texts contain period attitudes, prejudices, and terminology that may be offensive or harmful in modern contexts

Files In This Release

• shards/ — compressed training data shards
• shard_manifest.json — machine-readable shard summary and source totals

Citation

If you reference this dataset, cite it as:

TimeCapsuleLLM World English 1800-1875, sharded v1 corpus release.

If you need fuller provenance, use the accompanying source manifests and project records from the build environment.