Fine-News

Fine-News is a multilingual corpus of 635,068,343 news documents from the INFINI-NEWS Corpus. INFINI-NEWS extracts article text from Common Crawl News web archives. Fine-News applies language checks, quality filters, and duplicate removal to that text, and stores each retained document with a sampling weight. The corpus covers August 2016 through April 2026, using the month of the web crawl.

At a glance

Measure	Value
News documents	635,068,343
Sum of sampling weights	1,058,934,511
Capture months	117 (2016-08 through 2026-04)
Language-and-script labels	388
Parquet files	18,441
Compressed data size	1.03 TB

Each document occupies one physical row. Its sampling weight, from 1 through 10, specifies its relative sampling frequency when used for training. The earlier 2021–2025 corpus remains available at the v1 revision. V1 repeats rows for weighting and groups them by publication month.

Coverage

Documents per year

Years and months refer to the capture timestamp in the Web ARChive (WARC) record. A capture timestamp records when the crawler retrieved a page; the article's publication date comes from page metadata when available.

Capture year	Documents	Months
2016	3,571,834	August–December
2017	35,065,855	January–December
2018	45,012,132	January–December
2019	59,429,708	January–December
2020	88,594,020	January–December
2021	88,953,618	January–December
2022	87,715,960	January–December
2023	83,074,387	January–December
2024	67,478,657	January–December
2025	61,398,038	January–December
2026	14,774,134	January–April

Largest language groups

Language labels combine an ISO 639-3 language code with an ISO 15924 script code. The counts describe the classifier's assignments and leave classification accuracy unmeasured.

Language-and-script label	Documents
eng_Latn	219,192,987
spa_Latn	86,557,664
rus_Cyrl	46,936,249
ita_Latn	41,942,863
deu_Latn	40,346,858
fra_Latn	27,261,374
tur_Latn	23,184,136
arb_Arab	18,903,626
por_Latn	15,670,582
hin_Deva	11,000,537

The research/ directory contains coverage and retention tables for each processing stage, grouped by capture month, language, and publisher domain. research/coverage_summary.json records physical and weighted document totals.

Schema

Files use the path data/<capture-month>/<language-and-script>/part-*.parquet. The top-level columns are:

Column	Type	Description
text	string	Article text after preparation and quality filtering.
id	string	Source warc_record_id, retained as the document identifier.
media	list of structs	Optional DataTrove media field.
metadata	struct	Source fields, language evidence, duplicate-family fields, and sampling weight.

The metadata struct preserves the source URL, WARC filename, capture timestamp, payload digest, publication metadata, and extraction versions. The fields used for grouping and sampling are:

Field	Description
metadata.capture_month	UTC WARC capture month.
metadata.year_month	UTC WARC capture month.
metadata.warc_date	WARC capture timestamp.
metadata.publish_date	Publication date extracted from page metadata, when available.
metadata.language	Assigned language-and-script label.
metadata.minhash_cluster_size	Number of members in the verified duplicate family within its capture-month and language partition.
metadata.duplicate_group_id	Duplicate-family identifier.
metadata.sampling_weight	Adaptive training weight, from 1 through 10.

Quick start

The default Hugging Face view loads text, id, and metadata. Use PyArrow to read all four Parquet columns, including media.

Read one file in batches:

from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

path = hf_hub_download(
    repo_id="ksolovev/fine-news",
    filename="data/2026-04/eng_Latn/part-000000.parquet",
    repo_type="dataset",
    revision="v2",
)
parquet = pq.ParquetFile(path)
for batch in parquet.iter_batches(batch_size=4096):
    texts = batch.column("text").to_pylist()
    weights = batch.column("metadata").field("sampling_weight").to_pylist()
    # Apply the weights when sampling documents for training.

The dataset loader's train view contains all article files. Load one capture year with name="year_2024"; year subsets cover 2016 through 2026. Use metadata.sampling_weight when sampling documents for training.

Dataset creation

Source and preparation

The pipeline reads extracted Parquet from INFINI-NEWS and accepts 1,372,589,195 source rows. Preparation assigns the capture month, checks the extractor's language evidence, recovers raw HTML where possible, and removes rows that exceed tokenizer safety limits. The resulting source and processing counts are:

Processing step	Documents
Accepted source	1,372,589,195
Preparation	1,232,354,074
Duplicate-family representatives before quality selection	977,888,955
Quality-passing representatives	635,068,343
Final physical rows	635,068,343

Filtering and duplicate removal reduce the accepted source by 53.73% in physical rows.

Quality filtering

Quality filters evaluate every prepared occurrence before the pipeline selects duplicate-family representatives. The filters apply URL rules, boilerplate correction, Gopher repetition checks, FineWeb quality checks, and Gopher quality checks. Language-specific thresholds come from bundled FineWeb2 configurations. Repetition thresholds receive a 1.15 multiplier for news, and word counts must remain between 100 and 15,000.

Duplicate removal and weighting

MinHash generates candidate pairs within each capture-month and language partition. Verification groups identical full text and compares other candidates using Jaccard similarity between sets of word 5-grams, at a threshold of 0.90. Changed text at the same normalized URL and texts with different numeric sequences remain separate. Each removed member must directly match a retained representative; a chain of candidate matches cannot establish membership.

The pipeline selects a quality-passing occurrence as the representative, then uses the earliest capture timestamp and stable document ID to break ties. Families with no quality-passing occurrence are removed. Adaptive weights use family survival rates within each capture-month and language partition, while preserving every retained representative. rehydration.json records the weights and source counts.

The Fine-News code documents the method and reproduction commands. The source archive and processing configuration are available under research/provenance/.

Limitations

The 0.90 duplicate threshold is a heuristic based on lexical similarity. Lexically similar texts can differ in meaning or contain meaningful edits. The English language-score threshold is 0.80 and lacks corpus-level accuracy qualification. Some languages use fallback tokenizers.

Coverage, language classification, and filter retention vary across publishers, languages, and time. Retention counts describe the effects of processing; corpus quality and representativeness require separate evidence. Article text can contain personal information, errors, and harmful content.

Terms of use

Article text remains subject to its publishers' copyright and source-specific terms. Fine-News assigns no dataset-wide text license. Users must assess the rights and terms that apply to their intended use, including the INFINI-NEWS terms. Those terms restrict bulk article text to academic research. INFINI-NEWS releases its derived metadata under CC-BY-4.0. The Fine-News code uses AGPL-3.0-only.

Citation

If you use Fine-News, cite this dataset and the INFINI-NEWS paper:

@misc{solovev2026finenews,
  author = {Solovev, Kirill},
  title = {{Fine-News}},
  year = {2026},
  version = {2},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ksolovev/fine-news}
}

@misc{lazzaroni2026infininews,
  author = {Lazzaroni, Ruggero Marino and Lasser, Jana and Solovev, Kirill},
  title = {{Infini-News}: Efficiently Queryable Access to 1.3 Billion Processed Common Crawl News Articles},
  year = {2026},
  eprint = {2605.18337},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  doi = {10.48550/arXiv.2605.18337},
  url = {https://arxiv.org/abs/2605.18337}
}