frenchtext/bank-es-2401

Dataset Card for "bank es websites - 2401"

Dataset extracted from public websites by wordslab-webscraper in 2401:

• domain: bank
• language: es
• license: Apache 2.0

Dataset Sources

wordslab-webscraper follows the industry best practices for polite web scraping:

• clearly identifies itself as a known text indexing bot: "bingbot"
• doesn't try to hide the user IP address behind proxies
• doesn't try to circumvent bots protection solutions
• waits for a minimum delay between two pages to avoid generating too much load
• respects the website "robots.txt" indexing directives
• respects the web page Meta Robots HTML tag
• respects the web page X-Robots-Tag HTTP header
• respects the web page links rel=nofollow HTML attributes

The text data was extracted from the following websites:

Website	HTML pages	PDF docs	Words
elpais.com	648	0	859572
orangebank.es	230	2	83540
selectra.es	1781	14	2358891
www.20minutos.es	565	0	445259
www.bancamarch.es	134	74	391999
www.bancobig.es	83	18	56040
www.bancocooperativo.es	348	139	1140507
www.bancodepositos.es	112	111	429784
www.bancomediolanum.es	186	281	1631194
www.bancomundial.org	123	0	198375
www.bancosantander.es	514	0	426308
www.bankinter.com	1289	126	1170979
www.bbva.es	796	174	1159709
www.bcc.es	429	188	2213459
www.bde.es	147	0	75821
www.bnpparibas.es	107	15	187157
www.caixabank.es	331	7	275867
www.cetelem.es	277	12	191143
www.cnmv.es	304	164	763320
www.deutsche-bank.es	424	155	1278585
www.ebnbanco.com	314	10	226346
www.elperiodico.com	1997	0	1464925
www.evobanco.com	610	2	502912
www.finanzas.com	1389	0	612969
www.fundacionbancosabadell.com	482	33	237122
www.fundacionbancosantander.com	192	59	277875
www.grupbancsabadell.com	280	142	3587581
www.ibercaja.es	368	8	247437
www.lavanguardia.com	1417	0	1138809
www.openbank.es	97	0	89034
www.r4.com	5944	524	3347625
www.santander.com	165	48	186743
www.triodos.es	385	92	919441
www.unicajabanco.es	401	188	1004625

Uses

WARNING

• the text included in this dataset belongs to its original authors and is protected by copyright laws
• you are not allowed to use this dataset for anything else than training a large language model
• when using a large language model trained on this dataset, you will need to ensure that you comply with the law
• if you benefit from this large language model, you should try to share the value with the original text authors

wordslab-webscraper uses an advanced Html to text conversion algorithm optimized for long context language modeling:

• tries to recover the logical structure of the document from the Html or PDF layout
• preserves document / section / list / table grouping and nesting information
• deduplicates text at the website level while preserving the document structure

Each example in this dataset is a markdown text conversion of a full HTML page or PDF document:

• the document structure is preserved by markdown syntax: headers, lists, tables, paragraphs
• all duplicate paragraphs are removed

Dataset Structure

The dataset is divided in 3 splits:

• train: 80% of the data
• valid: 10% of the data
• test: 10% of the data

wordslab-webscraper generates one parquet file per website and per split.

The parquet files are named with the following pattern:

• bank_es_2401_[split]_[website].parquet

Note than you can load individual splits or websites with HuggingFace datasets using the following commands:

from datasets import load_dataset

# Load a single plit
dataset = load_dataset("namespace/bank-es-2401", split="train")

# Load a single website
data_files = { "train": "bank_es_2401_train_[website].parquet", "valid": "bank_es_2401_valid_[website].parquet", "test": "bank_es_2401_test_[website].parquet" }
dataset = load_dataset("namespace/bank-es-2401", data_files=data_files)

Each example in the dataset contains the text of a full web page or PDF document, with the following features:

• Uri: string
• ExtractedFromPDF: bool
• Timestamp: string
• Lang: string
• Title: string
• Text: string
• Words: int32
• AvgWordsLength: int32
• Chars: int32
• LetterChars: int32
• NumberChars: int32
• OtherChars: int32

Note that beause each example is a full page or document, the "Text" feature can be a pretty long string containing thousands of words (as measured by the "Words" feature): you will typically need to chunk it down to the context size of your large language model before using it.

Bias, Risks, and Limitations

This dataset is a direct extraction from the source websites.

It was not manually curated to remove misleading, offensive, or harmful content.

Please add a filtering step before using it to train a large language model if the source websites can't be trusted.

Dataset Card Contact

Please add a comment in the community section of this repository if you want the maintainer to add or remove websites from this dataset.