Open Fiction Corpus

A community-built, reproducible corpus of legally redistributable fiction for continued pretraining and genre-specific language-model research. Each row is one complete literary work with provenance and rights metadata — not pre-cut training chunks.

The catalogue, schemas, cleaning rules, and build code live in the GitHub repository: https://github.com/graystales-daina/open-fiction-corpus. This Hugging Face repository stores only generated release artifacts. Nothing here is hand-edited.

Status: the first versioned release is in preparation. This repository does not yet contain data files; they will appear here when the first release is published, and this card will be updated with the exact release tag and catalogue commit.

Dataset structure

Files

• data/books.jsonl.gz — every released work, one JSON object per line
• packs/<pack>.jsonl.gz — curated subsets (general fiction, genre packs) selected by the pack definitions in the source repository
• release/build-manifest.json — release version, catalogue commit, and the exact work IDs in every file
• release/checksums.sha256 — SHA-256 checksums of all data files

Row fields

Field	Description
id	Stable work identifier, e.g. se-wells-the-time-machine-en
title	Work title
authors	List of author names
language	BCP-47-style language code
form	novel, novella, short-story, or short-story-collection
primary_genre	Exactly one genre from the controlled vocabulary
genres	All applicable genres
subgenres	Subgenres from the controlled vocabulary
rights_status	Reviewed redistribution basis, e.g. public-domain
quality_status	Review state, e.g. human-reviewed
source_provider / source_identifier / source_revision	Pinned upstream source
text	The complete cleaned text of the work

The text contains no model-specific BOS/EOS strings and is not pre-tokenized. Training code should append the selected tokenizer's real EOS token and create context windows itself, without crossing document boundaries unless the framework supports document-aware packing.

Provenance and rights

Every work passes a per-work rights review before release. The catalogue records the upstream source and pinned revision, the asserted rights status, the jurisdictions actually reviewed, and the supporting evidence. Works with uncertain rights are excluded from builds by the tooling itself.

Important: public-domain status is jurisdiction-specific. A work may be public domain in the jurisdictions recorded in its rights review and still protected elsewhere. Translations and edited editions can carry independent copyright even when the original is public domain. See DATA_LICENSE.md. Dataset users remain responsible for evaluating their own use and location; this card is not legal advice.

Composition

No works have been released yet; the catalogue and first release are in preparation. Once published, this section will summarize the release's composition (work counts, forms, author balance, genre breakdown), and the exact work IDs per file will be listed in release/build-manifest.json.

Release packs restrict content.origin to human, so AI-generated works are excluded. Each work also records content.ai_assistance (none, proofreading, substantive, or unknown); packs do not currently filter on that field, so downstream users who need stricter guarantees should filter rows on it themselves.

Intended use

Causal-language-model continued pretraining on complete fiction, genre-specific fine-tuning, and prose-style research — especially for Small Language Models. See the training guide.

Out of scope: the corpus does not teach instruction following, dialogue-format chat, or outline-to-prose conversion by itself; those need separate task-tuning stages.

Limitations and biases

• Because rights review favours older works, the corpus over-represents pre-1930s prose, vocabulary, and social attitudes, including period racial language and colonial framing in some works. Works with such content are flagged in the catalogue rather than silently edited.
• Early releases skew toward British and US literature in English.
• OCR-derived sources may retain minor transcription errors; severe cases are excluded by quality flags.
• Corpus-balance targets (author caps, women-authored share, short-fiction share) are enforced per release and documented in the source repository's book backlog.

Reproducibility

Every release is built from a tagged version of the source repository, and this section always names the exact tag and commit the published artifacts were built from. No release has been published yet.

Rebuilding from the recorded catalogue commit with ofc build reproduces the release artifacts. The exact work IDs included in every file are listed in release/build-manifest.json.

Contributing and reporting problems

Nominate works, dispute classifications, or report text problems via GitHub issues: https://github.com/graystales-daina/open-fiction-corpus/issues. A work being in another dataset is not sufficient grounds for inclusion here.

Citation

@misc{openfictioncorpus2026,
  title  = {Open Fiction Corpus},
  author = {{Open Fiction Corpus contributors}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/open-fiction-corpus/corpus},
  note   = {Pre-release; release tag and catalogue commit will be recorded here}
}