Tralalabs/multilingual-news

A multilingual news text dataset collected from a wide range of publicly accessible news websites and normalized into a simple schema for language modeling, corpus analysis, and non-commercial research.

Dataset Summary

Tralalabs/multilingual-news is a multilingual corpus of news documents gathered from multiple online news sources across different countries and languages. The dataset is intended for research, experimentation, corpus analysis, multilingual NLP, and non-commercial model training/evaluation.

Each record contains:

• text: extracted article text
• url: source article URL
• language: language tag for the document

Main goals

• Provide a simple multilingual news corpus for experimentation
• Support multilingual language modeling and text processing
• Make it easier to build or test multilingual pipelines with a unified schema
• Offer a practical dataset for filtering, deduplication, language balancing, and corpus cleaning workflows

---

Dataset Structure

Data fields

Each row contains the following fields:

• text (string) The extracted article body text or the main cleaned article content.

• url (string) The original article URL used as the source reference for the document.

• language (string) The language tag assigned to the document, such as:

  • es-UR
  • es-MX
  • es-ES
  • pt-BR
  • en-US
  • etc.

Example row

{
  "text": "El gobierno anunció nuevas medidas económicas este martes...",
  "url": "https://example-news-site.com/article/12345",
  "language": "es-UR"
}

---

Supported format

The dataset may be distributed in Parquet format for efficient loading and streaming.

Example schema:

{
    "text": Value("string"),
    "url": Value("string"),
    "language": Value("string")
}

---

Languages

This dataset is multilingual and may contain documents from multiple language varieties and regions depending on the source collection process.

Examples of language tags that may appear include:

• en
• en-US
• en-GB
• es
• es-ES
• es-MX
• es-AR
• es-CL
• es-CO
• es-PE
• es-UR
• pt
• pt-BR
• fr
• de
• it
• nl
• and others depending on the included sources

The exact language distribution depends on the current dataset release.

---

Source Data

This dataset was created by collecting article text from a variety of publicly accessible news websites. It is an aggregated corpus rather than a single-source archive.

Important notes about source data

• Article content originates from third-party publishers.
• Copyright for original article text remains with the respective publishers/authors.
• URLs are included to preserve source traceability.
• Different sites may have different article structures, formatting conventions, and publication standards.
• Some articles may contain minor extraction artifacts depending on the source HTML structure.

---

Collection Process

At a high level, the dataset was created through a pipeline similar to the following:

1. URL discovery

   • Sitemap parsing where available
   • Public article link discovery
   • Source-specific crawling logic where needed

2. Article extraction

   • Extraction of main article text from page HTML
   • Removal of obvious boilerplate where possible
   • Basic normalization of whitespace and formatting

3. Schema normalization

   • Conversion into a shared format:

     • text
     • url
     • language

4. Filtering

   • Empty or very short documents removed
   • Basic deduplication and normalization
   • Language tagging or source-based language assignment

5. Export

   • Saved to Parquet for efficient storage and downstream loading

---

Intended Uses

Recommended uses

This dataset is intended for non-commercial uses such as:

• multilingual language modeling experiments
• corpus analysis
• tokenizer training experiments
• language identification experiments
• data cleaning / deduplication research
• domain adaptation experiments
• multilingual summarization research
• news-domain text classification research
• retrieval and indexing experiments
• educational and personal research projects

Potential use cases

• Pretraining or continued pretraining experiments
• Building multilingual news subsets by language
• Studying regional differences in journalistic language
• Evaluating multilingual normalization and filtering pipelines
• Testing data quality heuristics on real-world web text

---

Out-of-Scope / Discouraged Uses

This dataset is not intended as a legal replacement for publisher archives, and it should not be treated as a source of unrestricted commercial content.

Discouraged uses include:

• commercial redistribution of article text
• republishing large amounts of article content as a substitute for the original publishers
• use cases that ignore the copyright status of source materials
• uses that violate the terms of the original news websites

---

Data Quality Notes

Because the dataset is built from many websites, data quality can vary across documents. Possible issues include:

• duplicate or near-duplicate articles
• incomplete article extraction on some sites
• residual navigation text or boilerplate
• inconsistent punctuation/whitespace normalization
• different writing styles and formatting conventions
• regional language variation inside the same language family
• occasional mislabeled language tags if source tagging was imperfect
• short updates, live blog fragments, or syndicated content

Users are encouraged to run additional filtering or deduplication depending on their use case.

---

Bias, Risks, and Limitations

News-domain bias

This dataset reflects the editorial choices, biases, and coverage patterns of the included publishers. It should not be treated as a neutral representation of world events or public opinion.

Geographic and language imbalance

Some languages, countries, or publishers may be overrepresented relative to others. The dataset should not be assumed to be balanced across:

• regions
• political viewpoints
• topics
• publication dates
• article lengths
• languages

Publisher-specific conventions

Different outlets use different style guides, formatting, and article templates. This can create domain-specific artifacts in the text.

Temporal variation

News writing changes over time, and the corpus may mix documents from different years or periods depending on the crawl source and update process.

Copyright and provenance limitations

This dataset aggregates text originating from external publishers. Users are responsible for evaluating whether their intended use is compatible with the dataset license, the source websites’ terms, and applicable copyright law.

---

License

License for this dataset release: CC-BY-NC-SA-4.0

• BY: attribution required
• NC: non-commercial use only
• SA: derivatives/shared versions should be distributed under the same license

Copyright Notice

Important: The original article texts included in this dataset originate from third-party news publishers and authors. Copyright in the original articles remains with the respective rights holders.

This dataset card and release metadata are provided under CC-BY-NC-SA-4.0, but users should understand that the dataset aggregates material originating from external publishers. Users are responsible for ensuring that their use of the dataset complies with:

• applicable copyright law
• the terms of the original source websites
• any relevant downstream restrictions in their jurisdiction

If you are a rights holder and believe content should be removed from this dataset release, please open an issue or contact the dataset maintainer.

---

Attribution

If you use this dataset in research, please consider citing:

@misc{tralalabs_multilingual_news,
  title        = {Tralalabs/multilingual-news},
  author       = {Tralalabs},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Tralalabs/multilingual-news}}
}

If your use case relies heavily on a particular source subset, please also consider citing or crediting the original publisher(s) where appropriate.

---

Loading the dataset

Python / Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("Tralalabs/multilingual-news")
print(ds)
print(ds["train"][0])

Loading a specific split or shard

from datasets import load_dataset

ds = load_dataset("Tralalabs/multilingual-news", split="train")
print(ds[0]["language"])
print(ds[0]["url"])
print(ds[0]["text"][:500])

---

Example filtering by language

from datasets import load_dataset

ds = load_dataset("Tralalabs/multilingual-news", split="train")

spanish_uruguay = ds.filter(lambda x: x["language"] == "es-UR")
print(spanish_uruguay[0])

---

Example deduplication sketch

from datasets import load_dataset

ds = load_dataset("Tralalabs/multilingual-news", split="train")

seen = set()
clean_rows = []

for row in ds:
    key = (row["url"], row["text"][:500])
    if key in seen:
        continue
    seen.add(key)
    clean_rows.append(row)

print(len(clean_rows))

---

Recommended preprocessing before model training

Depending on your use case, you may want to apply additional steps such as:

• URL-level deduplication
• near-duplicate detection
• boilerplate removal
• minimum/maximum length filtering
• language verification
• profanity filtering if required for your project
• date range filtering
• source balancing
• country or language balancing
• removal of article updates/live-coverage fragments
• quality scoring / classifier-based filtering

---

Maintenance

This dataset may be updated over time as sources, extraction pipelines, or filtering rules change.

Possible future improvements:

• stronger deduplication
• per-source metadata
• publication dates
• titles
• source domain field
• better language validation
• quality scores / filtering annotations

---

Contact

Maintainer: Tralalabs

If you find extraction issues, malformed rows, duplicated content, or rights-related concerns, please open an issue in the relevant project/repository or contact the maintainer through the dataset page.

---

Disclaimer

This dataset is provided as is, without warranties of any kind. The maintainers make no guarantees about:

• completeness
• correctness
• legal suitability for a particular use case
• absence of copyrighted or duplicated material
• balance or representativeness across languages and regions

Use responsibly, verify your downstream pipeline, and review the licensing/copyright situation for your own jurisdiction and application.