agentlans/gutenberg-chunk-classification

Project Gutenberg Chunk Classification

A curated dataset containing random text chunks (blocks of contiguous lines) sampled from diverse Project Gutenberg *.txt files. Each chunk is automatically annotated using Qwen3.5 4B to determine its functional/literary category, language, text quality, and completeness.

This dataset is designed for research and training in text block segmentation, functional role classification (e.g., distinguishing narrative prose from dialogue, headings, footnotes, or poetry), and data quality filtering.

Dataset Structure

Data Fields

Each row in the dataset contains the following fields:

• id: Unique UUIDv4 identifier for the chunk.
• text: The raw text string (contiguous block of lines from a Gutenberg source).
• label: The functional/literary category of the text block.
• language: ISO 639-1 language code (e.g., en, de, fr) identified for the primary language.
• text_quality: Assessment of extraction cleanliness (clean, minor_corruption, severe_corruption).
• completeness: Indication of whether the chunk forms a complete unit (complete, fragment, uncertain).

Example Row

{
  "id": "e0457c87-c127-45ea-beaf-fffe15b4f46d",
  "text": "Dazu noch die Angaben in No. 1 (S. 67), No. 60, 62, 63, ferner No.\n66, 70, 98a, 100a und 105 (S. 130 und 132).",
  "label": "HEADING",
  "language": "de",
  "text_quality": "clean",
  "completeness": "complete"
}

Annotation Methodology

The dataset was annotated using Qwen3.5 4B enforcing strict JSON output constraints.

Click here for the complete prompt.

Classify the supplied text into exactly ONE category.

Identify the primary function and literary form of the text. Do not classify based solely on individual words, language, formatting, or sentence length.

## CATEGORIES

- **NARRATIVE**: Ordinary prose, including narration, descriptions, exposition, essays, arguments, historical accounts, scientific writing, and reflective writing.
- **DIALOGUE**: Conversations, spoken exchanges, interviews, transcripts, and dramatic speech presented as spoken language.
- **VERSE**: Poetry and intentionally lineated literary writing, including blank verse, free verse, metrical verse, verse drama, and poetic soliloquies. Rhyme is not required.
- **HEADING**: Chapter, section, subsection, and internal document headings.
- **TITLE**: Titles of books and other complete works, author names, and publication credits.
- **CAPTION**: Captions and labels associated with illustrations, photographs, figures, tables, and diagrams.
- **LIST**: Lists, enumerations, indexes, catalogues, dictionaries, glossaries, and bibliographic entries.
- **FOOTNOTE**: Footnotes, endnotes, source notes, and explanatory references.
- **INSTRUCTION**: Recipes, procedures, operating directions, and practical step-by-step guidance.
- **JUNK**: Unusable OCR, meaningless fragments, and extraction artifacts that cannot reasonably be assigned another category.

## CLASSIFICATION RULES

1. Classify according to the primary function of the text.
2. Language does not determine category. Apply the same categories to all languages.
3. Use NARRATIVE for ordinary continuous prose, including literary prose.
4. Use DIALOGUE for text primarily representing spoken communication.
5. Use VERSE for poetry and intentionally lineated literary compositions, including unrhymed poetry and blank verse.
6. A soliloquy written in verse is VERSE. A soliloquy written as ordinary spoken prose is DIALOGUE.
7. Quoted speech within ordinary narration does not automatically make the passage DIALOGUE.
8. A narrative poem is VERSE, not NARRATIVE.
9. Numbered entries and enumerations are generally LIST, not FOOTNOTE.
10. Distinguish TITLE from HEADING:
    - TITLE identifies a complete work or publication.
    - HEADING identifies a section within a work.
11. Use INSTRUCTION only when the text directs someone to perform actions or follow a procedure. Descriptions of actions are NARRATIVE.
12. Do not classify text as JUNK merely because it is short, incomplete, archaic, poorly formatted, or written in unfamiliar language.
13. Archaic spelling, historical vocabulary, and unusual grammar are not corruption.
14. When a passage contains multiple features, select the category representing its dominant function.
15. If uncertain, choose the most plausible category. Do not default to JUNK.

## TEXT QUALITY

Select exactly ONE:

- **clean**: Readable and substantially intact.
- **minor_corruption**: Mostly readable, with occasional OCR errors, missing characters, incorrect spacing, or minor extraction defects.
- **severe_corruption**: Extensive OCR errors, missing or garbled text, or extraction defects that substantially interfere with interpretation.

Do not confuse historical spelling or typography with corruption.

## COMPLETENESS

Select exactly ONE:

- **complete**: Forms an apparently complete unit, such as a paragraph, poem, heading, list entry, dialogue exchange, or note.
- **fragment**: Visibly incomplete, such as a sentence cut off mid-thought or an incomplete extraction.
- **uncertain**: Insufficient evidence to determine completeness.

A passage does not need to be a complete chapter or document to qualify as complete.

## LANGUAGE

Return the ISO 639-1 language code of the primary language (for example, "en", "fr", "de", "es").

Use "unknown" only when the language cannot be reliably identified.

For mixed-language text, identify the predominant language.

## OUTPUT

Return ONLY valid JSON.

Use exactly these five keys and no others:

{ "label": "NARRATIVE", "language": "en", "text_quality": "clean", "completeness": "complete" }

Do not provide explanations, confidence scores, additional fields, or Markdown formatting.

TEXT:

{{TEXT}}

Limitations & Considerations

• Local Context: Chunks are sampled independently without broader document context. Highly ambiguous passages (e.g., poetic prose or prose-like verse) can occasionally lead to classification ambiguity (such as distinguishing whether a dramatic soliloquy is classified as narrative, dialogue, or verse).
• Weak Labels: Labels are generated via model distillation (Qwen3.5 4B) and may contain occasional annotation artifacts or noise.
• Corpus Bias: Sourced exclusively from Project Gutenberg texts, meaning the language distribution leans heavily toward historical and classical literature.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.