README.md

---
license: odc-by
language:
- pt
pretty_name: PT-BR Blogs (long-form, C4-derived)
size_categories:
- 100K<n<1M
task_categories:
- text-generation
tags:
- pt-br
- brazilian-portuguese
- blogs
- long-form
- c4
- pretraining
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.parquet
---

# PT-BR Blogs (long-form, C4-derived)

Part of the **MagTina350m pretrain corpus release** by [Dataseek](https://dataseek.com.br)
under the Magestic.ai brand. This is one of nine silver-layer datasets that fed
[`dataseek/magtina350m-base`](https://huggingface.co/dataseek/magtina350m-base).

## Summary

185 K long-form Brazilian-Portuguese blog posts (≥ 5 K words each) extracted from C4 by filtering Blogspot, WordPress, Medium and similar platform domains. Higher per-document quality than generic web; useful for stylistic diversity and long-context training.

## Source and collection method

Source: AllenAI C4 PT-BR slice → blog-platform domain filter → minimum-length gate (≥ 5 000 words) → NFKD normalisation → URL dedup.

**ETL script (in the MagTina1B repository):** [`scripts/etl/14_blogs_c4_v1.py`](https://huggingface.co/dataseek/magtina350m-base) *(public release of the ETL scripts is on the roadmap; until then the data card below documents the recipe in full).*

## Filters and deduplication

The following filters were applied before this dataset reached its silver
(release-ready) state:

- Blog-platform domains only (blogspot, wordpress, medium, ghost, …)
- Word count ≥ 5 000 (long-form gate)
- URL dedup across web corpora

Global URL-normalised deduplication was applied across all web-derived corpora
(`webpages`, `news`, `blogs`) so the same article does not appear twice across
those three datasets.

## Schema

| Column | Type | Description |
|---|---|---|
| `id` | `int64` | Stable row id. |
| `text` | `string` | Blog post body. |
| `url` | `string` | Source URL. |
| `domain` | `string` | Hosting platform domain. |
| `timestamp` | `timestamp` | Crawl timestamp. |
| `n_chars` | `int32` | Character count of `text`. |

Columns dropped at export (kept private as ETL internals): `quality_score`

## Size statistics

| Metric | Value |
|---|---:|
| Rows | 185.0 K (185,012) |
| Characters | 6.92 B (6,923,317,314) |
| Estimated tokens (PT-BR, chars / 4.5) | 1.54 B |
| Compressed Parquet on disk | ~3.87 GB |

**Used in MagTina350m pretrain:** 1.540 B tokens
(8.9 % of MagTina350m's 17.39 B-token pretrain budget).

## How to load

```python
from datasets import load_dataset

ds = load_dataset("dataseek/ptbr-blogs", split="train", streaming=True)
for row in ds.take(5):
    print(row["text"][:200])
```

Streaming is recommended for the larger configs. For the smaller datasets
(`ptbr-dou`, `ptbr-books-publicos`) eager loading is fine.

## Licensing

ODC-BY 1.0 (CommonCrawl-derived). Long-form blog content; individual posts remain copyright of their authors. Use for non-commercial research is the safest interpretation; commercial use requires per-author clearance.

**Upstream attribution:** AllenAI C4 + individual blog authors

## Citation

If you use this dataset, please cite both the upstream source and MagTina350m:

```bibtex
@misc{magtina350m_pretrain_2026,
  title  = {MagTina350m pretrain corpus — PT-BR Blogs (long-form, C4-derived)},
  author = {Frasson, Ricardo and {Dataseek Team}},
  year   = 2026,
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/dataseek/ptbr-blogs}
}
```

Please also honour the upstream license terms — for CC-BY-derived data,
attribution to the upstream creators is mandatory; for CC-BY-SA, downstream
derivatives must remain CC-BY-SA-compatible.

## Intended use

- Pre-training, continued pre-training, or domain-adapting of Brazilian Portuguese
  language models.
- PT-BR NLP research where statistically representative public-web / academic /
  legal / encyclopedic data is needed.
- Reproducing or improving on the MagTina350m result.

## Known limitations and PII statement

- **Text was NOT PII-scrubbed.** URLs, emails, phone numbers and personal names
  that occurred in the source data may still be present. We strip zero-width
  characters and normalise Unicode but we do not run an NER pass.
- **Crawled data carries upstream biases** of CommonCrawl, Wikipedia, news outlets
  and academic institutions present in the source. We have not audited these.
- **No safety filtering** beyond langid and basic alpha-ratio gates. Hate-speech,
  spam and adult content present in the source remain unless caught incidentally.
- **Provenance preserved at row level.** Every row has either a `url`, `source` or
  `doc_id` column that points back to upstream — this is intentional, so consumers
  can re-license, redact or filter.

## Related releases

- **Model:** [`dataseek/magtina350m-base`](https://huggingface.co/dataseek/magtina350m-base) (354.6 M params, pretrained on this corpus + 8 sibling datasets)
- **Instruct model:** [`dataseek/magtina350m-instruct`](https://huggingface.co/dataseek/magtina350m-instruct)
- **Sibling datasets:** see `dataseek/ptbr-*` for all nine corpora

## License

[odc-by](https://spdx.org/licenses/odc-by.html)