rjzevallos/quebert_qu_bpe

RoBERTa for Quechua

This repository contains a RoBERTa-base model pretrained for the Quechua language using the Masked Language Modeling (MLM) objective. The model builds upon the corpus introduced in the QuBERT project and provides a RoBERTa alternative for downstream NLP research on Quechua and other low-resource Indigenous languages.

The goal of this work is to provide an open, high-quality pretrained language model that can serve as a strong foundation for a wide range of Natural Language Processing (NLP) tasks in Quechua.

---

Model Description

This model follows the RoBERTa-base architecture introduced by Liu et al. (2019) and was pretrained from scratch exclusively on Quechua text.

Model Details

Property	Value
Architecture	RoBERTa-base
Language	Quechua
Objective	Masked Language Modeling (MLM)
Tokenizer	Byte-Pair Encoding (BPE)
Framework	Hugging Face Transformers

---

Training Data

The model was pretrained on a curated monolingual Quechua corpus containing approximately 8 million words. The corpus was originally introduced in the QuBERT project and was compiled from multiple publicly available sources.

The training corpus includes text collected from:

• News articles
• Educational resources
• Government publications
• Religious texts
• Literature
• Wikipedia
• Blogs and other publicly available web resources

Before training, the corpus was carefully cleaned, normalized, and deduplicated to improve data quality.

The corpus primarily represents Southern Quechua, while also including material from other Quechua varieties whenever available.

More details about the corpus construction are available in the original QuBERT paper.

---

Training

The model was pretrained from scratch using the standard Masked Language Modeling (MLM) objective.

Unlike the original QuBERT model, which is based on the BERT architecture, this repository provides a RoBERTa-based language model trained on the same Quechua corpus.

---

Training Objective

Following the RoBERTa pretraining procedure, approximately 15% of the input tokens are selected for prediction.

The model learns contextual representations by reconstructing masked tokens from their surrounding context, enabling it to capture both syntactic and semantic information useful for downstream NLP applications.

---

Intended Uses

This model can be fine-tuned for a variety of downstream NLP tasks, including:

• Text Classification
• Named Entity Recognition (NER)
• Part-of-Speech Tagging (POS)
• Question Answering
• Semantic Similarity
• Information Retrieval
• Sentence Embeddings
• Document Classification
• Domain Adaptation through Continued Pretraining

Since the model is pretrained using MLM, it can also be used directly for masked token prediction.

---

Usage

Load the tokenizer and model

from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("rjzevallos/quebert_qu_bpe")
model = AutoModelForMaskedLM.from_pretrained("rjzevallos/quebert_qu_bpe")

Masked Language Modeling Example

from transformers import pipeline

fill_mask = pipeline(
    "fill-mask",
    model="rjzevallos/quebert_qu_bpe"
)

fill_mask("Payqa <mask> rinqa.")

---

Evaluation

The model can be evaluated using intrinsic and downstream metrics such as:

• Validation Loss
• Perplexity
• Named Entity Recognition (NER)
• Part-of-Speech Tagging (POS)
• Text Classification

Evaluation results can be added as they become available.

Metric	Value
Validation Loss	-
Perplexity	-

---

Limitations

Although the model has been trained on the largest publicly available monolingual Quechua corpus, several limitations remain:

• Performance depends on the dialect represented in the input text.
• Some Quechua varieties are underrepresented in the training corpus.
• Performance may decrease on specialized domains that are not well represented during pretraining.
• The model should not be used as the sole source for high-stakes or safety-critical decisions.

---

Bias and Ethical Considerations

Like all language models, this model may reflect linguistic and cultural biases present in the training corpus.

Researchers and practitioners are encouraged to evaluate the model carefully before deploying it in downstream applications, particularly those involving sensitive or high-impact use cases.

---

Citation

If you use this model in your research, please cite the original QuBERT paper:

@inproceedings{zevallos-etal-2022-introducing,
    title = {Introducing QuBERT: A Large Monolingual Corpus and BERT Model for Southern Quechua},
    author = {Zevallos, Rodolfo and
              Ortega, John and
              Chen, William and
              Castro, Richard and
              Bel, Núria and
              Yoshikawa, Cesar and
              Venturas, Renzo and
              Aradiel, Hilario and
              Melgarejo, Nelsi},
    booktitle = {Proceedings of the Third Workshop on Deep Learning for Low-Resource Natural Language Processing},
    year = {2022},
    publisher = {Association for Computational Linguistics},
    url = {https://aclanthology.org/2022.deeplo-1.1/},
    doi = {10.18653/v1/2022.deeplo-1.1}
}

---

Acknowledgments

This model builds upon the corpus introduced in the QuBERT project.

The original QuBERT work was partially funded by Project PID2019-104512GB-I00 from the Spanish Ministerio de Ciencia, Innovación y Universidades and the Agencia Estatal de Investigación.

We hope this model contributes to advancing Natural Language Processing research for Quechua and other low-resource Indigenous languages by providing an openly available pretrained RoBERTa model for the research community.