ptbr-scielo / README.md
dataseek's picture
Initial dataset release — MagTina350m pretrain corpus slice
a25f42f verified
---
license: cc-by-4.0
language:
- pt
pretty_name: PT-BR SciELO Articles (Brazilian Open-Access Research)
size_categories:
- 100K<n<1M
task_categories:
- text-generation
tags:
- pt-br
- brazilian-portuguese
- academic
- scientific
- scielo
- research
- pretraining
configs:
- config_name: default
data_files:
- split: train
path: data/*.parquet
---
# PT-BR SciELO Articles (Brazilian Open-Access Research)
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
154 K full-text Brazilian Portuguese research and review articles from SciELO Brazil (post-2010), spanning health sciences, social sciences, humanities and engineering. Avg ~34 KB per article — the largest academic-prose corpus in the MagTina350m mix.
## Source and collection method
Source: SciELO `articlemeta` API → fulltext HTML scrape → HTML→text → PT-only gate → year ≥ 2010 → doctype ∈ {research-article, review-article}.
**ETL script (in the MagTina1B repository):** [`scripts/etl/24_scielo_articles_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:
- doctype ∈ {research-article, review-article}
- year ≥ 2010
- lang = pt
- len(text) ≥ 500 chars
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 |
|---|---|---|
| `text` | `string` | Article fulltext. |
| `source` | `string` | Always 'scielo.br'. |
| `lang` | `string` | Language code (typically 'pt'). |
| `year` | `int32` | Publication year. |
| `doctype` | `string` | research-article | review-article. |
| `doc_id` | `string` | SciELO PID (links back to article). |
| `n_chars` | `int64` | Character count. |
Columns dropped at export (kept private as ETL internals): *none*
## Size statistics
| Metric | Value |
|---|---:|
| Rows | 154.2 K (154,218) |
| Characters | 5.29 B (5,291,892,295) |
| Estimated tokens (PT-BR, chars / 4.5) | 1.18 B |
| Compressed Parquet on disk | ~2.96 GB |
**Used in MagTina350m pretrain:** 1.176 B tokens
(6.8 % of MagTina350m's 17.39 B-token pretrain budget).
## How to load
```python
from datasets import load_dataset
ds = load_dataset("dataseek/ptbr-scielo", 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
CC-BY 4.0 — the dominant license across SciELO Brazil (open access). Individual articles may carry CC-BY-NC variants; downstream users should honour the per-article licenses available via the SciELO API. Attribution by article DOI is required for redistribution.
**Upstream attribution:** SciELO Brazil — https://scielo.br/
## 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 SciELO Articles (Brazilian Open-Access Research)},
author = {Frasson, Ricardo and {Dataseek Team}},
year = 2026,
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/dataseek/ptbr-scielo}
}
```
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
[cc-by-4.0](https://spdx.org/licenses/cc-by-4.0.html)