metadata

language:
  - ca

Dataset Card for Instrucat

Dataset Description

Dataset Summary

InstruCat is a dataset consisting of 216826 instructions in Catalan.

Dataset contains data converted to instructions format from the following datasets:

caBreu : The instructions were created in form of summarization tasks. There are 2 types of summarization categories in the dataset: extreme and abstractive. The extreme one summarizes text into one sentence and the abstractive into shorter texts around 3-5 sentences.
CatalanQA : The instructions correspond to questions in CatalanQA.
CaWikiTC : The instructions were created in 2 different ways of text classification tasks with the distribution 70% - 30%. The first way is to define a category of a given text. The second way is to answer where a given text belongs to a certain category in a form of alternative question.
ceil : The instructions were created in 2 different ways of Named Entity Recognition tasks with the distribution 70% - 30%. The first way is to list all the found Named Entities. The second way is to list only Named Entities of a particular category.
CoqCat : The instructions correspond to the first questions of CoqCat conversations.
GuiaCat : The instructions were created in form of sentiment analysis tasks.
IntoxiCat : The instructions were created in form of binary classification tasks. The task is to define wether a given text is toxic or no.
NLUCat : The instructions were created in form of phrase generation tasks to express a given intent.
Parafraseja : The instructions were created in form of text generation tasks. The task is to generate a text equivalent by meaning to a given text.
PAWS-ca : The instructions were created in form of text generation tasks. The task is to generate a text equivalent by meaning to a given text.
sts-ca : The instructions were created in form of text generation tasks. The task is to generate a text equivalent by meaning to a given text.
teca : The instructions were created in 2 different ways with the distribution 70% - 30%. The first way is in form of entailment generation tasks. The second way is to define whether one given text is an entailment of another given text.
WikiCat : The instructions were created in 2 different ways of text classification tasks with the distribution 70% - 30%. The first way is to define a category of a given text. The second way is to answer where a given text belongs to a certain category in a form of alternative question.

Dataset Structure

Data Splits

train.jsonl: 165100 instructions
validation.jsonl: 25351 instructions
test.jsonl: 26375 instructions

Data Instances

Three JSONL files, one for each split.

An example of 'test' looks as follows:

{
   "ID": "Parafraseja_8977",
   "instruction": "Reescriu aquesta frase sense alterar-ne el significat:",
   "context": "Es tracta d'un tipus que ens falla ja que a ell li falla aquesta falta d'interès per tal d'exercir el domini sobre l'ambient.",
   "response": "Es tracta d'un tipus que ens falla perquè a ell li falla aquesta falta d'interès per exercir el domini sobre l'ambient.",
   "category": "paraphrasis"
}

Category Distibution

Category	Number of instructions	%
ner	59410	27.39%
paraphrasis	34695	16.00%
text_classification	33393	15.40%
toxicity	29809	13.74%
qa	27427	12.64%
phrase_generation	11873	5.47%
entailment_generation	6354	2.93%
sentiment_analysis	5750	2.65%
abstractive_summarization	2999	1.38%
extreme_summarization	2999	1.38%
entailment	2117	0.97%

Acknowledgments

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215337, 2022/TL22/00215336, 2022/TL22/00215335 y 2022/TL22/00215334