---
license: cc0-1.0
task_categories:
  - text-generation
  - fill-mask
tags:
  - poetry
  - public-domain
  - byte-level
  - english
size_categories:
  - 1K<n<10K
---

# poetry-corpus

Clean public-domain English poetry, byte-level ready. Built for training
BDH (Baby Dragon Hatchling, vocab=256 raw bytes): every byte counts as one
token, stanza breaks preserved, poem boundaries marked.

## Files

| File | Contents |
|---|---|
| `poetry_corpus.txt` | Byte-level training rendering (`TITLE:` / `AUTHOR:` / body / `---`) |
| `poetry_corpus.jsonl` | One JSON record per line with metadata (schema below) |

## Schema (JSONL)

```
{"title": str, "author": str, "book": str, "gutenberg_id": int, "text": str}
```

Text rendering format (`.txt`):

```
TITLE: The Raven
AUTHOR: Edgar Allan Poe

Once upon a midnight dreary, while I pondered, weak and weary,
...
(stanza breaks preserved)
---
```

## Sources & licenses

54 books from Project Gutenberg (public domain), fetched via the
[gutendex](https://gutendex.com) API with a curated whitelist of poets who died
before 1931 (Shakespeare, Keats, Shelley, Blake, Byron, Tennyson, Browning,
Longfellow, Poe, Whitman, Dickinson, Emerson, Rossetti, Hopkins, Frost,
Yeats, Kipling, Stevenson, Carroll, Lear, Wilde, Wordsworth, Coleridge, ...).
Per-book manifest with Gutenberg IDs: `manifest.json` in the build repo.

## Cleaning

- Gutenberg boilerplate stripped (START/END markers, tolerant of line-wrap).
- Librivox audio-catalog pages detected and excluded at fetch time.
- Junk blocks (license remnants, `Language: English` metadata lines) and
  low-verse-ratio blocks (prose prefaces etc.) excluded.
- Poem splitting: stanza-based; title-like lines (short, no terminal
  punctuation, Title Case / CAPS / roman-numeral numbering) open a new poem.
  Books without clean structure are kept as whole-book records.
- Dedup: exact + near-dup word-3-gram Jaccard ≥ 0.85 across all books
  (anthology overlap is common).

## Known limitations

- Poem title detection is heuristic; numbered titles (`VI`, `XXXIII`) and the
  occasional editorial heading (`VARIANTS ON THE TEXT`) survive.
- A few prose prefaces pass the verse-ratio filter (short-line prose).
- Public-domain poetry skews pre-1930 by construction.

## Reproduction

Full pipeline + scripts: see the `create-datasets` project repo
(`scrapers/fetch_gutenberg_poetry.py`, `scripts/build_poetry_corpus.py`,
`qc/`, `CORPUS_REPORT.md`).