Access to this sample is granted automatically. Please share your name, company and email so we can follow up about The Web Index (TWI). Contact: contact@uricavelar.com

The Web Index (TWI) – Multilingual Parallel Sample

A free sample of The Web Index (TWI), a sentence/document-aligned multilingual knowledge corpus. This sample contains 1,000 aligned entries × 10 languages = 10,000 rows, with the same 32 columns as the full dataset.

Important: the text is AI-generated. English entries were written with large language models and then machine-translated into the other languages with LLMs. It is not human-written or human-translated, and it has not been fact-checked by humans. Treat it as synthetic data.

The full dataset (170 languages × 25,000 aligned entries, about 4.25 million rows) is available under a commercial or research licence. Contact: contact@uricavelar.com · thewebindex.ai

Maintainer: Uric Avelar · Dataset page: https://huggingface.co/datasets/thewebindexofficial/the-web-index-sample

Datasheet & licensing

Scope. Full release TWI-Clean v1 (October 2026): 170 languages (English source + 169 translations, 21 language families, dozens of low-resource languages) × 25,000 entries aligned 1:1 by meta_id = 4.25M rows, 32 columns, 196 categories, ~7 GB Parquet (one file per language + taxonomy.json + full changelog CSV).

Schema summary. Each row holds metadata (ID, language, category, slug), SEO title/description, the same content at three reading levels (lite / core / pro), 5 atomic RAG facts, 1 SFT instruction/answer pair, plus extras (IPA of the English title, 24×24 SVG icon, emoji, related-entry IDs). Full column list in Columns below.

Quality summary.

Structure checked on all 170 files: 25,000 rows each, identical meta_id order, 0% empty fields, no duplicate URLs.
Automated heuristic grades (untranslated text, wrong script, repetition loops, length outliers): 119 languages A, 35 B, 15 C. Languages graded D/F were removed.
≈49,500 logged corrections in the shipped changelog; about a dozen known residual issues are documented.
Limits: grades are heuristic ("A" = no visible defect, not a verified translation); no native-speaker or COMET evaluation yet.

AI-generated disclosure. All text is AI-generated: written in English by large language models, then translated by LLMs. No human authors, translators or fact-checking. No personal data or user-generated content, and nothing scraped from third-party sites. Use it as synthetic training / evaluation data, not as a source of facts.

Licences (USD, excl. taxes). Non-exclusive, perpetual for the delivered version, internal use, no redistribution of the raw data. Models trained on it may be used and distributed freely (commercial tier). English source included with every pack.

Tier	Content	Research / academic (non-commercial)	Commercial (incl. model training)
Free sample	1,000 entries × 10 languages (this repo)	CC BY-NC 4.0 (free)	—
Single language	25,000 entries + English	$150	$500
Language pack 10	Any 10 languages	$1,000	$3,000
Language pack 25	Any 25 languages or a regional pack	$2,000	$6,000
Full corpus	All 170 languages, 4.25M rows	$4,000	$12,000
Enterprise / custom	Full corpus + future versions, custom fields	On quote	On quote

The research tier is for universities, non-profits and public labs. Startups (<10 staff) may ask for a discount. Delivery is a secure download link within 2 business days of signature.

Contact: Uric Avelar · contact@uricavelar.com · thewebindex.ai

Languages in this sample

Code	Language	Heuristic quality grade (full file)
`en`	English	reference
`fr`	French	A (99.9)
`es`	Spanish	A (99.9)
`de`	German	A (99.7)
`zh-hans`	Chinese (Simplified)	A (96.4)
`ja`	Japanese	A (98.2)
`ar`	Arabic (MSA)	A (99.6)
`hi`	Hindi	A (99.8)
`sw`	Swahili	A (97.5)
`hne`	Chhattisgarhi	A (100.0)

Chhattisgarhi (hne) is included as an example of a low-resource language that received a top grade. Grades come from automated heuristics (untranslated text, wrong script, repetition loops, length outliers). An "A" means no visible defect was detected; it does not prove the translation is accurate.

Structure

One Parquet file per language in data/ (CSV copies are distributed separately). Every file has the same 1,000 meta_id values in the same order, so rows align across languages by meta_id.

from datasets import load_dataset
fr = load_dataset("thewebindexofficial/the-web-index-sample", "fr", split="train")

Columns (32)

Group	Columns	Description
Metadata	`meta_id`, `meta_lang_iso`, `meta_lang_dir`, `meta_category_id`, `meta_category`, `meta_slug`, `meta_url`, `meta_last_updated`	ID (alignment key), language code, text direction, category (196 categories), slug and original URL
SEO	`seo_title`, `seo_description`	Page title and meta description
Subject	`subject_anchor`, `subject_local`	English topic name and localized title
Reading levels	`core_body_1/2`, `lite_body_1/2`, `pro_body_1/2`	The same content at three registers: standard (core), simplified (lite), expert (pro)
RAG	`rag_point_1` … `rag_point_5`	Five short standalone factual statements
SFT	`sft_user_query`, `sft_assistant_answer`	One instruction/answer pair
Extras	`logic_visual_desc`, `logic_phonetic_ipa`, `logic_svg_geometry`, `logic_emoji_bridge`, `logic_machine_hex`, `logic_related_ids`	Visual description, IPA of the English title, 24×24 SVG icon, emoji code point, hash, 10 related entry IDs
Control	`end_guard`	Always `[END]`

Sampling

1,000 entries drawn at random (seed 42) from the 25,000 entries of the cleaned release (TWI-Clean v1), excluding 12 entries with known residual issues.

Intended uses

Evaluating or fine-tuning machine translation on multi-register parallel text
Multilingual RAG and question-answering prototypes
Readability / text simplification research (lite vs core vs pro)
Instruction-tuning experiments in low-resource languages

Limitations and biases

Synthetic content: may contain factual errors, hallucinations and stylistic artifacts typical of LLM output; translations may be literal or contain mistakes.
seo_description and some home/landing entries contain promotional text from the original website (e.g. article or language counts) that should not be taken as fact.
Quality grades are heuristic; no native-speaker or COMET evaluation has been done yet.
logic_phonetic_ipa is the IPA of the English title, not of the localized title.
Topics reflect the website's editorial taxonomy (internet culture, crypto, AI, flora and fauna, countries, languages, etc.) and are not a balanced encyclopedia.

Licence

This sample only is released under CC BY-NC 4.0: non-commercial use with attribution. Commercial use of the sample, and any use of the full dataset, requires a separate licence. Contact contact@uricavelar.com.

Citation

@misc{twi_sample_2026,
  author = {Uric Avelar},
  title  = {The Web Index (TWI): Multilingual Parallel Sample},
  year   = {2026},
  url    = {https://huggingface.co/datasets/thewebindexofficial/the-web-index-sample}
}

Downloads last month: 62

Total file size:

33.2 MB