Datasets:

Modalities:
Tabular
Text
Formats:
parquet
Languages:
Spanish
ArXiv:
Libraries:
Datasets
Dask
red_pajama_es_hq / README.md
ouhenio's picture
Update README.md
8073bef verified
metadata
language:
  - es
dataset_info:
  features:
    - name: text
      dtype: string
    - name: meta
      dtype: string
    - name: score
      dtype: float64
    - name: int_score
      dtype: int64
  splits:
    - name: train
      num_bytes: 1201679966776
      num_examples: 128920537
  download_size: 700567029628
  dataset_size: 1201679966776
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

RedPajama's High Quality Spanish subset

What is this?

The following is a high-quality dataset distilled from the Spanish subsection of RedPajama-Data-v2, created using the methodology proposed in FineWEB-Edu.

Usage

from datasets import load_dataset

ds = load_dataset("latam-gpt/red_pajama_es_hq") 

Filtering by quality score

Documents in this corpus are scored on academic quality from 2.5 to 5, with higher scores indicating better quality. The dataset can be filtered by score using standard filtering methods.

from datasets import load_dataset

ds = load_dataset("latam-gpt/red_pajama_es_hq")

# filter the dataset for scores > 3
filtered_ds = ds.filter(lambda x: x['score'] > 3)

Dataset creation

In a nutshell, we use Llama-3.1-70B to grade the educational quality of 550k samples from the original dataset. Then, we used these samples to train a encoder-based classifier, so that it learns to assign a score from 0 to 5. Since this model is cheaper to use than a GPT, we can run it at scale over the entire dataset, thus allowing us to filter a high-quality section from it.

Here is an overview of the architecture:

For more detailed information on how this dataset was created, refer to our open implementation.

What is Latam-GPT?

Latam-GPT is a Latin American initiative to develop a large language model built entirely in the region. The project encompasses all development stages — from data collection and pre-training to final model refinement — making it the first foundation model created completely within Latin America.

License

The text documents of the source database (RedPajama-Data-v2) were collected using 84 CommonCrawl snapshots, processed using the CCNet pipeline, and also provided under an Apache 2.0 license by the Together Computer team under the jurisdiction of the United States of America. There may be differences between the jurisdiction of the USA and Latin American countries. In order to comply with the terms of use of the Common Crawl Foundation and in the search for the greatest possible transparency, we provide the following contact to ask any questions, comments or complaints: eugenio.herrera@cenia.cl.