NoeFlandre/tiny-corneille

Tiny Corneille

Tiny Corneille is a clean French text corpus. It holds the seven tomes of the works of Pierre Corneille, from Project Gutenberg. The corpus follows the idea of Andrej Karpathy's tiny Shakespeare.

• Code, rules and reproduction: github.com/NoeFlandre/tiny-corneille

Stats

Measure	Value
Tomes	7
Characters	3,795,595
Lines	118,402
Distinct characters	119
GPT-2 tokens	1,571,834
Characters per GPT-2 token	2.415

GPT-2 tokens use tiktoken's gpt2 encoding.

Files

• tiny_corneille.txt: the corpus. UTF-8, LF line endings.
• manifest.json: the SHA-256 of each raw tome and of the corpus, and the expected statistics.

Sources

The tomes come from the edition of Charles Joseph Marty-Laveaux, as published by Project Gutenberg.

Tome	Gutenberg id
1	31628
2	34445
3	36011
4	39739
5	46527
6	48590
7	51612

Processing

The rules keep Corneille's words and remove the rest:

• Removed: the Gutenberg header and footer, the editor's front matter, editor notices, lists of collated editions, variant complements, editorial footnotes and their references, tables of contents, printed verse line numbers, and italic markers.
• Kept: the plays with their stage directions and character lists, Corneille's own prefaces and discourses (including the AVERTISSEMENTS of tome 1), and his examinations.
• Original spelling is kept.
• Known residue: some editorial text remains, such as bracketed variant verses and a few editor paragraphs. The data notes list it.

The full rules are in the data notes.

Reproduce

git clone https://github.com/NoeFlandre/tiny-corneille
cd tiny-corneille
uv sync --locked
uv run tiny-corneille fetch
uv run tiny-corneille build
uv run tiny-corneille verify

fetch checks each raw download against manifest.json. verify checks the built corpus byte for byte.

Licence

The works of Corneille are in the public domain. Project Gutenberg publishes the ebooks under its own terms, including its trademark notice. Keep that notice when you redistribute this data.