br-llm-data/wikibooks-instruct

Wikibooks Instruct pt-BR

Synthetic instruction-tuning dataset in Brazilian Portuguese generated from cleaned Portuguese Wikibooks text.

The published run contains all accepted examples from 20260630T233835Z-wikibooks-instruct-7fec6755.

Summary

• Accepted examples: 106,010
• Attempted examples: 248,040
• Original target: 127,330 examples
• Run status in metadata: partial
• Source documents selected: 9,095
• Split: train only
• Source dataset: costadev00/wikibooks-pt-br-extract
• Source revision: 6a64293ff7bb35dba45afe06b2063abc91790bb3
• Generator: google/gemma-4-31B-it-qat-w4a16-ct via local vLLM
• Evaluation: deterministic composite filters; no LLM judge coverage

Configs

Use the config that matches your training pipeline:

• messages: chat-style records with messages.
• alpaca: records with instruction, input, and output.
• prompt_completion: records with prompt and completion.
• canonical: richest internal export with all major fields.
• source_documents: selected source text snapshot used by the run.

Each instruction example includes provenance fields such as document_id, source_title, task, difficulty, evidence, generator, and model.

Task Mix

The run covers 14 task families:

• classification
• closed_qa
• comparison
• concept_explanation
• definition
• didactic_explanation
• fact_checking
• information_extraction
• rewrite
• short_answer
• structured_extraction
• summarization
• taxonomy
• timeline

Difficulty distribution:

• easy: 26,493
• medium: 53,093
• hard: 26,424

License and Attribution

This dataset is derived from Portuguese Wikibooks content via costadev00/wikibooks-pt-br-extract, which is licensed as cc-by-sa-3.0 on the Hub. Preserve appropriate attribution and share-alike obligations when redistributing derived versions.

Reproducibility Metadata

Execution metadata is included under metadata/:

• config.resolved.json
• manifest.json
• plan.json
• progress.json
• report.json
• corpus-index.jsonl

The internal SQLite run database is intentionally not included.