---
language:
- pt
language_creators:
- found
annotations_creators:
- no-annotation
multilinguality:
- monolingual
pretty_name: Pessoana
license: cc-by-sa-4.0
task_categories:
- text-generation
size_categories:
- 10M<n<100M
tags:
- text
- portuguese
- portuguese-literature
- literature
- prose
- poetry
- public-domain
- character-level
- eca-de-queiros
- machado-de-assis
- fernando-pessoa
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.txt
  - split: validation
    path: data/validation.txt
  - split: test
    path: data/test.txt
- config_name: full
  data_files:
  - split: train
    path: data/full.txt
---

# Pessoana

Pessoana is a Portuguese literary corpus of 84 public-domain works of Portuguese and
Brazilian classic literature, prepared for character-level language modeling. It is a
companion to [camoniana](https://huggingface.co/datasets/vreabernardo/camoniana), named
in homage to Fernando Pessoa (whose *Mensagem* it includes), and spans the canon from
Eça de Queirós and Machado de Assis to Florbela Espanca and Mário de Sá-Carneiro.

## Dataset Details

- **Language:** Portuguese (`pt`) — pre-1990 orthography.
- **Authorial focus:** Portuguese and Brazilian classic literature (84 works).
- **Modality:** text (prose and poetry).
- **Primary task:** text generation / character-level language modeling.
- **Annotations:** none.
- **Tokenizer assumption:** one Unicode character is one token.

## Files

```text
README.md
metadata.json
vocab.json
data/full.txt
data/train.txt
data/validation.txt
data/test.txt
```

- `data/full.txt`: full corpus text (train + validation + test).
- `data/train.txt` / `data/validation.txt` / `data/test.txt`: the splits.
- `metadata.json`: corpus statistics and split metadata.
- `vocab.json`: character vocabulary and integer mappings for character-level modeling.

## Dataset Structure

The default configuration exposes the train, validation, and test splits:

```python
from datasets import load_dataset

ds = load_dataset("vreabernardo/pessoana")
```

To load the full corpus as one split:

```python
from datasets import load_dataset

full = load_dataset("vreabernardo/pessoana", "full")
```

For raw-file workflows, download `data/full.txt` directly.

## Splits and Statistics

```text
works: 84
vocab size: 165
train characters: 15,044,011
validation characters: 1,848,444
test characters: 2,269,090
full characters: 19,161,545
split: document-level, overlap-audited (no work appears in more than one split)
```

Splits are assigned per document and overlap-audited, so no work leaks between train,
validation, and test. This keeps validation and test as honest held-out measurements
for character-level training runs.

## Intended Uses

- Training small character-level language models for Portuguese.
- Testing Portuguese text generation experiments.
- Reproducing character-level dataset preparation.

## Out-of-Scope Uses

- Treating the validation or test split as a benchmark for unrelated Portuguese
  language modeling systems.
- Assuming the corpus represents modern Portuguese usage.
- Using generated outputs as factual or scholarly editions of the source texts.

## Data Creation

The works were parsed from public-domain EPUBs to UTF-8 plain text and assigned to
train/validation/test at the document level with an overlap audit. Publisher front
matter — colophons (*Ficha Técnica*), tables of contents, ISBN/edition notices, and
licence statements — was stripped so the corpus is continuous literary prose and verse.
No labels or annotations were added.

## Limitations and Biases

This is a literary corpus. It reflects the language, spelling, style, and editorial
choices of its source material (mostly 19th- and early-20th-century texts) and should
not be treated as representative of contemporary Portuguese. Light extraction artifacts
(e.g. footnote markers) may remain.

## License and Provenance

The texts come from **Projecto Adamastor**, which publishes Portuguese-language literary
works under a Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)
licence. The underlying works are in the public domain by age. This compilation is
redistributed under CC BY-SA 4.0 — please keep attribution to Projecto Adamastor and the
original authors.

## Citation

Use the exact Hugging Face revision you used. Replace `<revision-sha>` with the commit
hash shown on the dataset page.

### APA

```text
Vrea, B. (2026). Pessoana: A Portuguese literary corpus for character-level
language modeling [Data set]. Hugging Face.
https://huggingface.co/datasets/vreabernardo/pessoana
Revision: <revision-sha>
```

### BibTeX

```bibtex
@misc{vrea2026pessoana,
  title     = {Pessoana: A Portuguese Literary Corpus for Character-Level Language Modeling},
  author    = {Vrea, Bernardo},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/vreabernardo/pessoana},
  note      = {Revision: <revision-sha>}
}
```