You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset is a curated, production-grade Swiss web corpus from OptiTransfer. The sample (10,000 records) is available upon approval for evaluation. The full dataset (110,491 records across 22 production files including Parquet, JSONL, language splits, and RAG chunks) is available under commercial licence -- delivered instantly via HuggingFace upon payment settlement (bank transfer, TWINT, or crypto). Contact data@optitransfer.ch for pricing. Please complete all fields to request access.

Log in or Sign Up to review the conditions and access this dataset content.

*.ch Swiss Web Premium (A+)

Grade A+ -- 112.4M tokens -- 110,491 records -- 29 languages -- Full provenance -- PII-redacted -- RAG-ready -- SFT-formatted

A production-grade Swiss web corpus from the .ch TLD namespace. 110,491 documents independently quality-scored (avg 93.3/100, minimum 90), PII-redacted, and SHA256-verified. Built for LLM training, RAG pipelines, SFT fine-tuning, and multilingual NLP.


OptiTransferData Portfolio

Premium sovereign web corpora for LLM training, RAG pipelines, and multilingual NLP. EU AI Act compliant, PII-redacted, SHA256-verified.

Dataset Records Type Repositories
*.ch Swiss Web Premium (A+) 110,491 docs, 78 fields Web corpus + SFT + RAG Sample / Storefront -- Full Dataset

Contact data@optitransfer.ch to discuss your data requirements or get notified when new datasets launch.


Dataset Summary

Property Value
Product Name *.ch Swiss Web Premium (A+)
Dataset ID cache_ch_ac906c8b
Total Records 110,491 unique source URLs
Total Tokens 112,428,245 (tiktoken cl100k_base / GPT-4)
Total Characters 367.5M
Total Words 51,785,182
Quality Grade A+ (avg 93.3/100, floor 90, ceiling 96)
Languages 29 (67.1% German, 20.0% French, 7.9% English, 4.4% Italian + 25 more)
Content Categories 13 (general, news, ecommerce, documentation, academic, legal, education, government, discussion, finance, encyclopedia, code, blog)
Unique Domains 19,927 (domain cap: 500 records per domain)
Crawl Source Common Crawl CC-MAIN-2026-12
Crawl Window 2013-05-18 to 2026-03-17
Total Size 554.1 MB (all formats combined)
RAG Chunks 196,304 pre-built (512-token windows, 64-token overlap)
Provenance Full WARC-level audit trail (filename, offset, digest)
Compliance EU AI Act, FADP-aligned, Swiss-hosted processing
Slop Score 0.001 mean (99.9% human-authored content)
Pipeline OptiTransfer Tier 6 Sidecar sprint_20_v6
Certificate OT-CERT-CACHE_CH
QA Report 43-page v6 PDF (full enumeration, cross-validation, provenance)

What Is in This Repository

Evaluation Sample (10,000 records -- available on approval)

  • data/train-00000-of-00001.parquet -- Stratified sample preserving language, category, and quality distribution (directly loadable via HuggingFace Datasets library)
  • docs/schema.json -- Full JSON Schema for all 78 fields
  • docs/data_card.md -- Detailed methodology and quality documentation
  • docs/manifest.json -- Complete dataset statistics and pipeline configuration
  • docs/quality_certificate.html -- Visual quality certificate
  • docs/domain_breakdown.csv -- All 19,927 domains with record counts
  • docs/skill_distribution.csv -- ML skill tag and difficulty distributions
  • docs/sample.csv -- Tabular sample summary
  • docs/SHA256SUMS -- SHA256 checksums for all 22 production files
  • reports/OptiTransfer_QA_Report_v6_cache_ch_ac906c8b.pdf -- 43-page QA report (v6)

Full Dataset (available under commercial licence)

The full dataset is available at OptiTransferData/swiss-web-premium-ch-full:

  • 7x Parquet shards (140.3 MB, zstd compressed) -- optimal for Spark, Arrow, BigQuery
  • 7x JSONL.gz shards (147.0 MB compressed) -- universal compatibility
  • 4x Language splits: DE (95.4 MB), FR (28.1 MB), EN (10.3 MB), IT (5.3 MB)
  • 4x RAG chunk shards (127.7 MB) -- 196,304 retrieval-ready chunks
  • SHA256SUMS for all 22 production files
  • Full 43-page QA report (v6)

Contact data@optitransfer.ch for full dataset pricing and licensing.


Quality Scoring Methodology

Every document is evaluated using a 9-component weighted quality algorithm. Only records scoring 90/100 or above are included in this A+ build.

Quality Score Distribution

Score Records Percentage
96 26,266 23.8%
95 15,365 13.9%
94 6,239 5.6%
93 20,401 18.5%
92 14,603 13.2%
91 15,089 13.7%
90 12,528 11.3%

Quality Component Means

Component Mean Score What It Measures
whitespace 99.95 Normalisation and formatting consistency
digit_ratio 99.02 Numeric-to-alphabetic character balance
repetition 98.44 Absence of boilerplate or repetitive passages
capitalization 98.25 Correct and consistent capitalisation
length 97.31 Document length relative to optimal range
language_confidence 97.09 FastText detection confidence
sentence_structure 96.87 Grammatical completeness and variety
punctuation 93.47 Punctuation density and correctness
paragraphs 49.51 Paragraph count and structural depth

Language Distribution

Language Records Percentage Dedicated Split
German (de) 74,130 67.1% Yes
French (fr) 22,072 20.0% Yes
English (en) 8,695 7.9% Yes
Italian (it) 4,823 4.4% Yes
Portuguese (pt) 187 0.2% --
Spanish (es) 164 0.1% --
Russian (ru) 98 0.1% --
Tamil (ta) 68 0.1% --
+ 21 more 254 0.2% --

All records originate from .ch TLD domains. Language distribution reflects the multilingual nature of Switzerland's web presence, with all four national languages represented.


Record Schema (78 fields)

Each record contains rich metadata for maximum ML utility.

Core Fields

Field Type Description
url string Source URL (unique, deduplicated)
text string Clean extracted text (PII redacted where detected)
language string ISO 639-1 language code (FastText lid.176)
language_confidence float Detection confidence (0-1)
quality_score int 9-component quality score (90-96 in this build)
quality_components object Individual scores: length, sentence_structure, paragraphs, capitalization, punctuation, digit_ratio, whitespace, repetition, language_confidence
token_count int GPT-4 token count (tiktoken cl100k_base)
word_count int Word count
sentence_count int Sentence count
perplexity float Character-level entropy (mean 3.02, IQR 0.14)
text_length int Character count
paragraph_count int Number of paragraphs
avg_sentence_length float Mean sentence length in words

Classification and Trust

Field Type Description
content_category string One of: general, news, ecommerce, documentation, academic, legal, education, government, discussion, finance, encyclopedia, code, blog
category_info object Semantic categorisation with confidence score and structural features
trust_tier string Domain trust level: established (70.4%), general (24.9%), news_editorial (1.9%), academic (1.7%), institutional (1.1%)
trust_info object Multi-factor trust scoring with component breakdown
domain string Source domain
tld string Top-level domain (always ch)
url_path string URL path component

ML-Ready Enrichments

Field Type Description
skill_tags array 31 distinct skill families with sub-variants (e.g., reasoning_arithmetic, coding_function_calling, creative_writing_fiction)
instruction_style string One of: detailed (65.1%), brief (31.2%), json_structured (1.3%), step-by-step (1.1%), conversational (1.0%), formal (0.2%)
difficulty_level string intermediate (70.7%) or expert (29.3%)
slop_score float AI-generated text marker (0 = clean human, 1 = heavy slop). Mean: 0.001
spatial_awareness bool Whether text contains meaningful spatial/positional descriptions (2.4% of records)
persuasion_audience string or null Target audience when persuasive content detected: commercial, public, academic, institutional

Provenance and Compliance

Field Type Description
timestamp string Original crawl timestamp (ISO 8601)
crawl_date string Crawl date
warc_source_key string WARC source key
mime string MIME type from WARC header
status string HTTP status code
digest string SHA1 content digest from WARC
filename string Source WARC filename (enables source reconstruction)
offset string Byte offset in WARC
length string Record length in WARC

Safety and PII

Field Type Description
pii_detected array PII types found: email, phone_intl, phone_swiss, credit_card, iban, ahv_swiss, ssn_us
pii_redacted_count int Number of token-level [REDACTED] replacements applied
repetition object Word repeat ratio, n-gram repeat ratio, line repeat ratio

Content-Specific (populated when applicable)

Field Type Description
code_language string or null Programming language (SQL, Python, JavaScript, etc.)
code_confidence float or null Code detection confidence
academic object or null arXiv ID, DOI, abstract, categories
news object or null Headline, publish date, source, category

Token and Perplexity Statistics

Token Distribution

Range Records Percentage ML Use Case
0-100 651 0.6% Short texts, classification training
100-500 39,760 36.0% SFT / RLHF examples, RAG chunks
500-1,000 33,072 29.9% Continued pre-training, summarisation
1,000-2,000 24,368 22.1% Long-form SFT, document understanding
2,000-5,000 11,159 10.1% Extended context, legal/encyclopedia
5,000+ 1,481 1.3% Long-context models

Stats: Mean 1,018 tokens/doc -- Median 680 -- Min 30 -- Max 17,200

Perplexity

Metric Value
Mean 3.02
Median 3.01
IQR 0.14 (2.94-3.08)
Min-Max 1.10-5.08

The tight IQR of 0.14 confirms consistent text quality across the entire corpus.


PII and Safety

All records have been processed through OptiTransfer's PII detection and redaction pipeline:

Metric Value
Records with PII detected 33,803 (30.6%)
Total redactions applied 88,643
PII types detected Email, phone (international), phone (Swiss), credit card, IBAN, AHV (Swiss), SSN (US)
Compliance GDPR-aligned, FADP-aligned, EU AI Act

All detected PII instances have been replaced with [REDACTED] tokens prior to dataset publication.


Content Categories

Category Records Percentage
general 72,752 65.8%
news 21,263 19.2%
ecommerce 8,112 7.3%
documentation 3,340 3.0%
academic 1,157 1.0%
legal 1,093 1.0%
education 919 0.8%
government 834 0.8%
discussion 596 0.5%
finance 248 0.2%
encyclopedia 99 0.1%
code 77 0.1%

Processing Pipeline

 1. WARC Extraction     Raw extraction from Common Crawl (CC-MAIN-2026-12)
 2. Swiss Transfer      Zero-copy AWS S3 to Exoscale SOS (Zurich, Switzerland)
 3. Text Extraction     Trafilatura HTML extraction (boilerplate removal)
 4. Language Detection   FastText lid.176 (176 languages)
 5. Quality Scoring     9-component weighted scoring (0-100 scale)
 6. Tokenization        tiktoken cl100k_base (GPT-4 compatible)
 7. Classification      Content categorisation + multi-factor trust scoring
 8. Metadata Extract    Code detection (15 languages), academic, news metadata
 9. PII Redaction       Email, phone, IBAN, credit card, AHV/SSN patterns
10. Domain Gate         Denylist filtering for known problematic domains
11. Deduplication       SHA1 content digest dedup + URL dedup
12. Quality Filter      Minimum quality score: 90
13. Output Generation   JSONL.gz sharded + Parquet (zstd) + RAG chunks + language splits

Filtering summary: 354,964 source records -> 110,491 delivered (31.1% pass rate)

Stage Records Removed
Quality filter 217,771
Deduplication 15,441
Domain cap (500/domain) 10,661
Low trust 94
Premium gate 82
URL dedup 125

Quick Start

Load the Sample with HuggingFace Datasets

from datasets import load_dataset

ds = load_dataset("OptiTransferData/swiss-web-premium-ch")
print(ds["train"][0])

Filter by Quality and Language

# Load and filter for high-quality German records
ds = load_dataset("OptiTransferData/swiss-web-premium-ch")
df = ds["train"].to_pandas()

high_quality_de = df[(df.quality_score >= 95) & (df.language == "de")]
print(f"{len(high_quality_de)} premium German records")

Load Full Dataset from Parquet (licensed customers)

import pandas as pd

# Load all 7 shards
dfs = [pd.read_parquet(f"cache_ch_ac906c8b_{i:03d}.parquet") for i in range(1, 8)]
df = pd.concat(dfs, ignore_index=True)
print(f"{len(df)} records, {df.token_count.sum():,} tokens")

Load Full Dataset via HuggingFace Datasets (licensed customers)

from datasets import load_dataset

ds = load_dataset("OptiTransferData/swiss-web-premium-ch-full")
print(f"{len(ds['train'])} records")
print(ds["train"][0])

Load from JSONL

import gzip, json

with gzip.open("cache_ch_ac906c8b_001.jsonl.gz", "rt") as f:
    for line in f:
        record = json.loads(line)
        print(record["url"], record["quality_score"], record["token_count"])

Use Language Splits

import gzip, json

# Load only French records
with gzip.open("cache_ch_ac906c8b_fr.jsonl.gz", "rt") as f:
    fr_records = [json.loads(line) for line in f]
print(f"{len(fr_records)} French records")

Use RAG Chunks

import gzip, json

with gzip.open("cache_ch_ac906c8b_rag_001.jsonl.gz", "rt") as f:
    chunks = [json.loads(line) for line in f]
print(f"{len(chunks)} RAG chunks (512-token, 64-overlap)")

Pricing and Access

Tier Contents Availability
Sample 10,000 records (this repository) Free -- request access above
Full Dataset 110,491 records + Parquet + JSONL + 4 language splits + RAG chunks (22 files) Commercial licence
Enterprise Multi-TLD packs, custom pipelines, priority support Custom pricing

How to Purchase the Full Dataset

  1. Evaluate -- Request access above and explore the 10,000-record sample
  2. Contact us -- Email data@optitransfer.ch to discuss licensing and receive a quote
  3. Settle payment -- We accept the following payment methods:
Method Details
Bank Transfer (SEPA / SWIFT) Invoice provided with full banking details upon agreement
TWINT Swiss instant payment -- settle directly via TWINT (details provided on invoice)
Crypto (BTC / ETH / SOL) Wallet addresses provided on request -- contact us for details
  1. Instant delivery -- Upon payment confirmation, you receive access to the full dataset on HuggingFace within 24 hours. All 22 production files (Parquet, JSONL, language splits, RAG chunks) delivered through HuggingFace secure infrastructure.

Enterprise and volume buyers: Custom invoicing, purchase orders, and multi-dataset bundles available. Contact us for tailored terms.

data@optitransfer.ch -- optitransfer.ch

OptiTransfer delivers sovereign, compliance-ready web corpora for AI teams building in regulated markets. Every dataset ships with full provenance, SHA256 verification, and Swiss-hosted processing.


Quality Assurance

This dataset is accompanied by a comprehensive 43-page QA report (v6) containing:

  • Full enumeration of all 110,491 records
  • Cross-validation across all supporting documents (10/11 pass, 1 finding within tolerance)
  • Language distribution analysis (29 languages)
  • Quality component breakdown (9-axis scoring)
  • Domain diversity analysis (19,927 unique domains)
  • PII detection and redaction audit (88,643 redactions)
  • Temporal coverage analysis (13-year span: 2013-2026)
  • Slop score analysis (AI content detection)
  • RAG suitability assessment (196,304 chunks)
  • Data provenance and traceability chain
  • Regulatory compliance statement (EU AI Act, FADP)

The report is available at reports/OptiTransfer_QA_Report_v6_cache_ch_ac906c8b.pdf in both the sample and full repositories.


Licence and Attribution

  • Source data: Common Crawl Foundation Terms of Use (commoncrawl.org/terms-of-use)
  • Processing and packaging: OptiTransfer.ch proprietary pipeline (Tier 6 Sidecar sprint_20_v6)
  • Redistribution: Subject to OptiTransfer.ch Terms of Service
  • Important: Page content may be copyrighted by original authors; use subject to applicable copyright laws.

Citation

@dataset{optitransfer_ch_2026,
  title     = {*.ch Swiss Web Premium (A+)},
  author    = {OptiTransfer},
  year      = {2026},
  url       = {https://huggingface.co/datasets/OptiTransferData/swiss-web-premium-ch},
  note      = {110,491 records, 112.4M tokens, Grade A+, .ch TLD, EU AI Act-aligned}
}

Product ID: cache_ch_ac906c8b -- Source: CC-MAIN-2026-12 -- Generated: 2026-03-24 -- Pipeline: OptiTransfer.ch Tier 6 Sidecar sprint_20_v6 -- Certificate: OT-CERT-CACHE_CH

Downloads last month
-