costadev00/wikipedia-pt-br-instruct-5k

Wikipedia PT-BR Instruct

wikipedia-pt-br-instruct is a synthetic supervised fine-tuning (SFT) dataset in Brazilian Portuguese generated from Wikipedia-derived documents.

This release is an intermediate evaluation dataset produced with the sft-dataset-creator pipeline from the run wiki-ptbr-extract-calib-5kdocs-14tasks. It was generated from a fixed revision of costadev00/wikipedia-pt-br-extract:

cdbd07dc4a3de6e64632c718710b3ae0ebaeb0ff

The dataset is intended for intermediate SFT experiments, qualitative evaluation and validation of synthetic-data generation workflows in Brazilian Portuguese. It should not yet be treated as a final, fully human-audited publication-grade corpus.

Dataset Summary

Field	Value
Source documents selected	5,000
Planned generation slots	70,000
Accepted examples before cleanup	53,705
Published examples after cleanup	53,704
Splits	train only
Language	Brazilian Portuguese
Generator	google/gemma-4-31B-it-qat-w4a16-ct
Evaluator	Deterministic filters only

One accepted example was removed before publication because it contained visible English leakage in the instruction text. The removal is documented in metadata/publish_manifest.json.

Available Configurations

The same examples are provided in three equivalent SFT formats.

messages

Chat-style format:

from datasets import load_dataset

ds = load_dataset(
    "costadev00/wikipedia-pt-br-instruct",
    "messages",
    split="train",
)

Each row contains:

• id
• messages: list with user and assistant turns
• source
• document_id
• source_title
• task
• difficulty
• evidence
• generator
• model

prompt_completion

Prompt/completion format:

from datasets import load_dataset

ds = load_dataset(
    "costadev00/wikipedia-pt-br-instruct",
    "prompt_completion",
    split="train",
)

Each row contains:

• id
• prompt
• completion
• source
• document_id
• source_title
• task
• difficulty
• evidence
• generator
• model

alpaca

Instruction/input/output format:

from datasets import load_dataset

ds = load_dataset(
    "costadev00/wikipedia-pt-br-instruct",
    "alpaca",
    split="train",
)

Each row contains:

• id
• instruction
• input
• output
• source
• document_id
• source_title
• task
• difficulty
• evidence
• generator
• model

Generation Method

The dataset was generated with a deterministic planning and filtering pipeline. The generator model proposes structured JSON candidates, while acceptance is decided by deterministic checks.

Pipeline overview:

1. Stream documents from costadev00/wikipedia-pt-br-extract.
2. Apply the wikipedia_ptbr eligibility profile.
3. Chunk selected documents into bounded context windows.
4. Plan 14 generation slots per selected document.
5. Generate candidates with a single local instruction model.
6. Parse and validate candidate JSON.
7. Accept or reject candidates using deterministic filters.
8. Write accepted examples to checkpoint shards and final exports.

The 14 planned task types were:

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

Difficulty was sampled with weights:

• easy: 0.25
• medium: 0.50
• hard: 0.25

Filtering Rules

Candidates were accepted only when they passed deterministic gates:

• valid JSON and expected schema;
• non-empty instruction and output;
• minimum instruction length;
• no visible references to hidden source text, such as "according to the text";
• no normalized duplicate among accepted examples;
• valid evidence spans pointing back to the source document chunk.

The main evaluator did not use a second LLM judge. This makes the acceptance decision auditable and reproducible from the run artifacts, but it does not guarantee full semantic correctness of every generated statement.

Audit

An AI-assisted audit was performed before publication. The audit verdict was:

Approved as an intermediate evaluation dataset with caveats.

Global checks before publication:

• Exported rows before cleanup: 53,705
• Published rows after cleanup: 53,704
• Normalized duplicate examples in the Alpaca export: 0 detected
• Hidden-source references in visible fields: 0 detected
• Obvious English leakage after cleanup: 0 known cases

The audit report is included in:

metadata/codex_verdict.md

The publish manifest is included in:

metadata/publish_manifest.json

Intended Use

Recommended uses:

• intermediate SFT experiments in Brazilian Portuguese;
• qualitative inspection of synthetic instruction data;
• evaluation of the sft-dataset-creator generation and filtering pipeline;
• ablation studies comparing SFT formats and task mixtures.

Not recommended as-is for:

• final benchmark publication without additional human audit;
• high-stakes factual evaluation;
• legal, medical or financial instruction tuning;
• claims that require fully verified semantic grounding.

Limitations

The deterministic filters verify format, provenance, deduplication and evidence span validity, but they do not prove that every sentence in every answer is fully supported by the source text.

Known limitations:

• Some examples may be low utility because they are derived from sparse or category-like Wikipedia text.
• Some task labels can be semantically loose.
• The duplicate filter is exact after text normalization and does not catch all semantic paraphrases.
• Evidence spans indicate recoverable support, but not necessarily complete answer-level entailment.
• The run ended as partial: 53,705 accepted examples were produced from 70,000 planned slots before cleanup.

Provenance

Each example includes:

• source dataset name;
• source document ID;
• source title;
• task type;
• difficulty;
• evidence spans;
• generator backend and model.

Additional metadata files:

• metadata/report.json: run-level metrics;
• metadata/config.resolved.json: resolved generation configuration;
• metadata/generation_info.json: export metadata;
• metadata/publish_manifest.json: publication cleanup manifest;
• metadata/codex_verdict.md: AI-assisted audit verdict.

License And Redistribution

The source corpus is derived from Wikipedia material. This dataset card keeps the repository license as other until redistribution terms and attribution requirements are reviewed for a final public release.

Users should verify licensing and attribution requirements before redistribution or production use.

Citation

If you use this dataset, please cite the dataset repository and the generation pipeline used to create it.

@dataset{wikipedia_pt_br_instruct_2026,
  title = {Wikipedia PT-BR Instruct},
  author = {costadev00},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/costadev00/wikipedia-pt-br-instruct}
}