AVIM-AI/gemma-4-E4B-es-latam-mlx
Text Generation • 7B • Updated • 865
Datasets:
You must accept the conditions to access this dataset
This repository is openly available, but you must accept the conditions of use and share your contact information with the repository administrators to access its files and content.
By selecting “Accept conditions and access repository,” you agree to the
following conditions:
Each document retains the license or condition recorded in its metadata.
The user is responsible for reviewing and complying with the applicable
attribution, reuse, and redistribution requirements.Entities with annual revenue exceeding USD 1 billion, or that have raised
cumulative capital exceeding USD 1 billion, require CENIA's prior written
authorization before accessing, using, or redistributing
LatamGPT-Corpus-1.0, referred to in these conditions as the LatamCorpus,
or any part of it. This restriction does not apply to academic research,
development by public institutions or non-profit civil-society
organizations, or specialized applications built on pre-existing
foundation models. See "Restriction on use by large enterprises" below
for full details.The corpus is provided “as is.” CENIA and the contributing institutions
do not warrant that the content is complete, accurate, current, or fit
for a particular purpose.The user assumes responsibility for the use of the corpus, its copies,
transformations, derived datasets, and models developed from it.Some datasets may contain technical traceability signals, such as
statistical watermarks, designed to assess whether they were used in
training an artificial-intelligence model.CENIA may suspend access when an entity with annual revenue exceeding
USD 1 billion, or that has raised cumulative capital exceeding USD 1
billion, accesses, uses, or redistributes the LatamCorpus without
CENIA's prior written authorization.
By accepting, you also authorize the sharing of your name, username, and
email address with the repository administrators for the purpose of
maintaining an access record for the corpus.
Log in or Sign Up to review the conditions and access this dataset content.
LatamGPT-Corpus-1.0
🌐 Language versions: English | Español | Português
🔗 Project links: Official LatamGPT website | Corpus dashboard
🤖 Associated model: The complete LatamGPT corpus—of which this repository contains the openly released portion—was used in the training process of Llama-3.1-70B-LatamGPT-SFT-1.0.
Dataset description
Summary
LatamGPT-Corpus-1.0 is the open release of the data corpus assembled for the continued pretraining of LatamGPT, a language model developed by CENIA together with a regional network of partner institutions.
The dataset contains text in Spanish, Portuguese, and English associated with Latin America and the Caribbean. It was built to strengthen the representation of the region's language varieties, institutions, cultures, and knowledge domains in large-scale language-model training.
This repository includes only sources authorized for open release under the LatamGPT data-governance framework.
Scope of the statistics
The complete LatamGPT-Corpus-1.0 pretraining corpus contains approximately 296.5 billion tokens. Unless explicitly stated otherwise, distributions by language, country, topic, source, and dataset refer to the complete corpus, while file counts, record counts, and storage size refer to the open Hugging Face release.
Detailed project statistics and dataset-level breakdowns are available in the LatamGPT corpus dashboard.
Dataset information
| Property | Value |
|---|---|
| Curated by | LatamGPT team (CENIA and partner institutions) |
| Languages represented in the open release | Spanish, Portuguese, and English |
| License | Multiple licenses by source; see Licensing |
| Project dashboard | Complete-corpus statistics and dataset breakdowns |
| Release scope | Sources authorized for open release |
| Associated model | https://huggingface.co/latam-gpt/Llama-3.1-70B-LatamGPT-SFT-1.0 |
| Contact | latam-gpt@cenia.cl |
Uses
⚠️ Restriction on use by large enterprises
Regardless of the openness level described above, access, use, or redistribution of any data released by the project by a large enterprise is strictly conditioned on obtaining prior written permission from CENIA. Failure to comply with this policy will be considered a violation of the governance terms.
For the purposes of this section, a large enterprise is understood to be any entity that, considered together with its parent companies, subsidiaries, and other related entities, records annual revenue exceeding USD 1 billion or has raised cumulative capital exceeding USD 1 billion, considering capital contributions, capital-convertible instruments, and comparable financing. Entities that do not exceed either of these thresholds may use the corpus for the purposes described in this section without prior permission. These entities are invited to inform CENIA of their use for traceability purposes, without this constituting a condition of access. This condition is evaluated as of the date the corpus is obtained; exceeding the threshold afterward does not affect models already trained with data obtained previously. CENIA may update these thresholds in future versions of the corpus.
For the purposes of this restriction, a foundation model is understood as an artificial-intelligence system trained on large volumes of data, of general purpose, and adaptable to a wide variety of downstream tasks through fine-tuning, adaptation, or programming interfaces. Commercial use is understood as use directed at the economic exploitation of the resulting model or of services derived from it, whether directly or indirectly, regardless of whether the model is distributed openly or in a closed manner.
This restriction does not apply to academic research, development by public institutions or non-profit civil-society organizations, or the development of specialized applications or models built on top of pre-existing foundation models. Under no circumstances do these exceptions authorize the use of the data by a large enterprise to train foundation models intended for commercial use.
Rationale: reciprocity
This condition responds to a principle of reciprocity. LatamGPT-Corpus was built with the voluntary contribution of public, academic, and civil-society institutions across Latin America and the Caribbean, through a collective effort aimed at strengthening the region's technological capabilities. Its release seeks to empower that community, not to subsidize the development of commercial models whose benefits do not return to the ecosystem that made the corpus possible.
CENIA acknowledges that this condition means that LatamGPT-Corpus, considered as a compilation, does not qualify as open source under the Open Source Initiative's definition nor as open data under the Open Definition, since both exclude restrictions based on field of use or type of user. This is a deliberate decision, adopted after weighing the trade-off between unrestricted openness and the sustainability of the regional effort. CENIA will review this condition in future versions of the corpus in light of accumulated experience.
Scope and enforcement mechanisms
This restriction operates through two complementary channels:
a. Repository access terms. Access to the repositories where CENIA makes LatamGPT-Corpus available is subject to prior acceptance of these access terms, which the user expressly accepts at the time of requesting or performing the download. Downloads from the repositories, including those at the Open Access Level, will be subject to a technical gated-access mechanism requiring express acceptance of these terms as a prior step. Obtaining the data from these repositories constitutes acceptance of these conditions and creates a contractual relationship between the user and CENIA, governed by Chilean law and submitted to the jurisdiction of the ordinary courts of justice seated in the commune of Santiago, Chile.
b. Condition on the compilation. Independently of the above, this restriction applies to LatamGPT-Corpus as an integrated compilation curated and distributed by CENIA, comprising its selection, organization, processing, enrichment, metadata, and structure — elements that constitute the project's own intellectual contribution. This condition does not modify the original licenses applicable to third-party data incorporated into the corpus, nor does it limit their use when obtained independently from their original sources. Anyone who redistributes LatamGPT-Corpus as a compilation, in whole or in substantial part, must incorporate these same terms as a condition of the redistribution and record them in the corresponding documentation or metadata.
Consequences of non-compliance
Non-compliance with this restriction constitutes a violation of the access terms and of LatamGPT-Corpus's governance and licensing conditions, and may result in revocation of access, suspension of granted authorizations, and the exercise of any legal actions applicable under relevant law.
Authorization requests
Organizations that require permission under this section may direct their request to latam-gpt@cenia.cl. CENIA will evaluate each case considering, among other factors, the initiative's contribution to the Latin American and Caribbean AI ecosystem.
Note on traceability mechanisms: released datasets into which this mechanism has been incorporated contain technical traceability signals, such as statistical watermarks, designed to assess whether they were used to train an artificial-intelligence model.
Note on compliance: each dataset preserves its original license in the metadata. End users assume responsibility for complying with those licenses, while the project maintains a permanent channel to identify and remove any data that lacks a compatible license (contact:
latam-gpt@cenia.cl).
Direct use
This dataset is intended for:
- Continued pretraining (CPT) or domain-adaptive pretraining of language models.
- Research on the linguistic representation of Latin America.
- Studies of regional varieties and multilingual corpora.
- Corpus analysis and dataset-curation research.
Dataset structure
The corpus is distributed in 101 Parquet shards:
train-00000-of-00101.parquet
...
train-00100-of-00101.parquet
After accepting the conditions on the repository page, authenticate your machine:
hf auth login
The dataset can be loaded with 🤗 Datasets:
from datasets import load_dataset
dataset = load_dataset(
"latam-gpt/LatamGPT-Corpus-1.0",
split="train",
streaming=True,
)
example = next(iter(dataset))
print(example)
Streaming is recommended for initial exploration because the complete open release is approximately 385 GB. To download and materialize all shards locally, omit streaming=True and ensure that sufficient disk space is available.
Data fields
| Field | Type | Nulls | Description |
|---|---|---|---|
id |
string | 0% | Unique document hash |
text |
string | 0% | Document text |
language |
string | 0% | Detected language (spanish, portuguese, english) |
text_length |
int64 | 0% | Length of text in characters |
license |
string | 0% | License associated with the original source |
country |
string | ~6.11% | One or more geographic labels separated by commas; the annotation space covers 20 countries and territories |
topic |
string | ~26.36% | One or more topic labels separated by commas; the taxonomy covers ten main categories |
dataset_name |
string | 0% | Dataset identifier |
Data splits
| Split | Examples | Shards | Size |
|---|---|---|---|
| train | 143,402,544 | 101 | ~384 GB |
These figures refer exclusively to the files included in the open Hugging Face release.
Example
{
"id": "b82686ade44fe302e85754ef7f1679c9",
"text": "Relatos de mi pueblo: Una obra clásica de la mano de Efraín M. Pacheco García, leyendas, ritos, habladurías. en el que une tradiciones y costumbres de los pueblos mexicanos.Efrain M. Pacheco García nos ofrece una docena de relatos breves (cuentos y leyendas) para los niños. Están ambientados en Hecelchakán (Campeche-México)...",
"language": "spanish",
"text_length": 896,
"license": "Open Data",
"country": "Mexico",
"topic": "Arts, Humanities & Social Sciences, Indigenous Peoples",
"dataset_name": "latamgpt_7"
}
Complete-corpus distributions
Scope of the figures
The charts in this section describe the complete LatamGPT-Corpus-1.0 pretraining corpus, containing approximately 296.5 billion tokens, and not exclusively the open Hugging Face release.
Country and topic annotations are multilabel: a document may contribute to more than one category. The percentages are therefore non-exclusive and are not expected to sum to 100%.
The categories and their distributions are shown below and are available in greater detail in the project dashboard.
Topic token coverage
The complete corpus is organized around ten high-level topic labels covering political, social, scientific, educational, cultural, economic, health, media, sports, and Indigenous knowledge domains.
Figure 1. Multilabel token coverage by topic in the complete LatamGPT-Corpus-1.0 pretraining corpus.
Politics and Humanities & Social Sciences have the largest label-level token coverage in the complete corpus, followed by Economy & Finance, Education, Communication & Media, and Exact Sciences. Medicine & Health, Sports & Recreation, and Indigenous Peoples have lower coverage in this corpus version.
Country token coverage
The geographic annotation system covers 20 Latin American and Caribbean countries and territories.
Figure 2. Multilabel token coverage by country in the complete LatamGPT-Corpus-1.0 pretraining corpus.
Argentina and Mexico have the largest label-level token coverage, followed by Colombia, Chile, and Brazil. The observed distribution reflects the availability and composition of the collected sources and should not be interpreted as demographic weighting, population weighting, or an intentionally uniform allocation across countries.
Interactive charts and dataset-level details are available in the LatamGPT corpus dashboard.
Licensing
This dataset aggregates documents from multiple collections released under different licenses.
The license applicable to each document is recorded in the license field.
This release corresponds to the Open Access Level (Orange) of the LatamGPT source catalog and includes documents under licenses or conditions compatible with open release, including CC0, CC BY variants, Open Data, ODC-By, and Apache 2.0.
The following document-level license distribution applies exclusively to the open Hugging Face release. The counts sum to the 143,402,544 records distributed in this repository.
| License | Documents | Share |
|---|---|---|
| Apache 2.0 | 74,592,182 | 52.02% |
| CC0 1.0 | 54,239,524 | 37.82% |
| ODC-By v1.0 | 11,764,895 | 8.20% |
| CC-BY-4.0 | 2,406,640 | 1.68% |
| Open Data | 312,635 | 0.22% |
| CC BY 2.5 AR | 66,818 | 0.05% |
| CC BY 3.0 BR | 11,863 | <0.01% |
| CC BY 2.5 PE | 5,557 | <0.01% |
| CC BY 2.0 | 2,430 | <0.01% |
Users are responsible for complying with the license associated with each document.
Traceability mechanisms
Note on traceability mechanisms: released datasets that incorporate this mechanism contain technical traceability signals, such as statistical watermarks, designed to assess whether they were used in training an artificial-intelligence model.
Personal and sensitive information
The corpus contains documents from public, institutional, and web sources.
During processing, automated anonymization filters and measures aimed at reducing the presence of personal and sensitive information were applied in accordance with the project's data-protection criteria.
These mechanisms should not be considered exhaustive. The corpus may contain identifiable information present in the original sources.
Users should apply any additional review and filtering measures appropriate to their use case.
Biases, risks, and limitations
The corpus reflects the characteristics, availability, and limitations of its original sources. Therefore:
- Representation across languages, countries, and topics is not uniform.
- Quality, currency, accuracy, and licensing may vary across documents.
- Offensive, biased, outdated, inaccurate, or identifiable content may be present.
- Automated curation and filtering processes reduce risks but do not guarantee their complete removal.
Users should assess whether the dataset is appropriate for their use case and apply any additional filtering, review, and validation measures that may be required.
Citation
@misc{latamgptcorpus2026,
title = {LatamGPT-Corpus-1.0},
author = {LatamGPT Team},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/latam-gpt/LatamGPT-Corpus-1.0}}
}
Acknowledgments
LatamGPT is made possible by the collaborative work of technical teams, institutions, communities, and organizations contributing to the development of open and representative artificial intelligence for Latin America and the Caribbean.
- Downloads last month
- 393