# Dataset Card for Pho-BP

## Dataset summary

Pho-BP is a multitask Brazilian Portuguese benchmark with 46,000 examples covering rhyme detection, syllable segmentation, initial-phoneme identification, and reading-error classification. The release uses fixed splits, explicit phonetic annotations, executable evaluation scripts, and machine-readable baseline reports.

## Supported tasks

1. **Rhyme detection:** binary classification over word pairs.
2. **Syllable segmentation:** prediction of inter-character syllable boundaries.
3. **Initial phoneme identification:** 24-class classification.
4. **Reading-error classification:** four-class imbalanced classification.

## Dataset structure

The repository exposes four Hugging Face configurations. Every configuration has `train`, `validation`, and `test` splits. See `README.md` for the full column schemas and loading examples.

## Data generation and curation

### Lexical source

Words were selected from `fserb/pt-br`, an MIT-licensed Brazilian Portuguese lexical collection. Its ICF values were used to prioritize more frequent lexical items. Filtering retained NFC-normalized lowercase alphabetic forms between 2 and 18 characters.

### Phonological processing

IPA transcriptions were produced by a deterministic Brazilian Portuguese rule-based G2P procedure included in the release-generation protocol. Rhyme keys are the segment from the final primary stress marker through the word ending. Syllabification uses Pyphen's `pt_BR` patterns and is represented both as hyphenated syllables and binary inter-character boundary labels.

### Split policy

Rhyme and syllable splits use deterministic SHA-256-based partitioning before example selection. Initial-phoneme examples are sampled per class and then divided without word overlap. Reading-error target words are separated by deterministic split keys. The release seed is `20260714`.

### Reading-error privacy policy

No child identifier, recording, demographic attribute, school, location, session identifier, or free-text note is present. Because the original row-level child corpus was not provided, public examples are generated through transparent linguistic transformation rules and marked `privacy_preserving_rule_based_public_release`.

## Dataset statistics

- Total examples: 46,000
- Rhyme label distribution: 50/50 in every split
- Syllable mean word length: 8.70
- Initial-phoneme classes: 24
- Reading-error distribution: 45/25/20/10 percent

## Baseline results

The JSON files reproduce the thesis table values and carry provenance metadata stating that they are reported results from the original benchmark. They should be treated as documentary baselines, not as scores recomputed on this public reconstruction.

## Biases and limitations

The lexical source contains specialist and low-frequency forms despite ICF ranking. Automated G2P and hyphenation may encode tool-specific conventions and may not capture Brazil's full dialectal variation. Rule-generated reading errors cannot reproduce timing, prosody, self-correction, contextual effects, or learner-level distributions from authentic oral reading.

## Responsible use

Use this dataset for aggregate model research and benchmarking. Do not infer a learner's competence, disability, diagnosis, or educational placement from model predictions. Authentic learner-data extensions require informed consent, governance approval, de-identification, and a separate data-access protocol.

## Licensing

MIT; see `LICENSE`.