napolab 1.0.0

Natural Portuguese Language Benchmark

Natural Portuguese Language Benchmark (Napolab)

The Natural Portuguese Language Benchmark (Napolab) is designed to be a collection of Portuguese datasets that are:

• Natural: Composed, as far as possible, of natural Portuguese text or text that has undergone professional translation.
• Reliable: The standard metrics used to evaluate the tasks on the dataset should reliably correlate with human judgments (such as accuracy, F1 score, Pearson correlation coefficient, etc.).
• Public: All datasets must be available and downloadable via a publicly accessible link.
• Human: All datasets must be annotated by human experts without any automation.
• General: The datasets should not necessitate domain-specific knowledge or any preparation for an educated Portuguese speaker to solve the suggested tasks.

This repository contains links to demos, models fine-tuned on the benchmark, and instructions for using the datasets in the most convenient manner. The benchmark currently includes the following datasets:

• assin
• assin2
• rerelem
• hatebr
• reli-sa
• faquad-nli
• porsimplessent

We are open to expanding the benchmark, and suggestions for future additions are welcome in the issues. We also welcome evaluation results from any models on this benchmark, and we are particularly curious about the outcomes of evaluating recent LLMs on these datasets.

In addition to making these datasets available in Portuguese, all datasets have also been automatically translated into English, Spanish, Galician, and Catalan using the facebook/nllb-200-1.3B model through the Easy-Translate library.

The small napolab.py script in this repository includes a convenient load_napolab_benchmark function that will download the entire Napolab benchmark from the Hugging Face Hub, including all translated versions:

from napolab import load_napolab_benchmark

napolab = load_napolab_benchmark(include_translations=True)

benchmark = napolab["datasets"]
translated_benchmark = napolab["translations"]

Demos

All of our fine-tuned models have been integrated into an appropriate Hugging Face Space.

Interact with our models in your browser by exploring our demos:

• Portuguese Offensive Language Detection
• Portuguese Question Answering
• Portuguese Semantic Similarity
• Portuguese Textual Entailment
• Portuguese Text Simplification

Models

Summary of the Fine-Tuning procedure

• Step 1. Hyperparameter optimization is performed using quasi-random search based on Google's Deep Learning Playbook instructions. The best learning rate, weight decay, and adam beta1 parameters for each Transformer model on each task are identified.

• Step 2. The best hyperparameters from step 1 are used to fine-tune each model 40 times with different random seeds for up to one epoch ( Dodge et al. (2020) ). The 10 best models after the first epoch are selected for the next step.

• Step 3. The top 10 models from step 2 are fine-tuned for 20 epochs, generating predictions for the test set ( Mosbach et al. (2021) ). We select the model that is closest to the average of predictions (for regression tasks) or the mode of predictions (for classification tasks). This final model is then uploaded to the Hugging Face Hub and displayed in the tables below.

Links and Results

Our fine-tuning procedure has achieved results that are either slightly superior or at the same level as the previous state-of-the-art (if any). Below is a summary of the results achieved on each dataset.

ASSIN 2 - STS ( Semantic Textual Similarity )

Model	Pearson	MSE
ruanchaves/bert-large-portuguese-cased-assin2-similarity	0.86	0.48
Previous SOTA ( for Pearson ) - Souza et al. (2020)	0.852	0.50
SOTA ( for MSE ) - Stilingue	0.817	0.47
ruanchaves/mdeberta-v3-base-assin2-similarity	0.847	0.62
ruanchaves/bert-base-portuguese-cased-assin2-similarity	0.843	0.54

ASSIN 2 - RTE ( Recognizing Textual Entailment )

Model	Accuracy	F1
ruanchaves/bert-large-portuguese-cased-assin2-entailment	0.90	0.90
ruanchaves/mdeberta-v3-base-assin2-entailment	0.90	0.90
Previous SOTA	0.90	0.90
ruanchaves/bert-base-portuguese-cased-assin2-entailment	0.88	0.88

ASSIN - STS ( Semantic Textual Similarity )

Model	Pearson	MSE
ruanchaves/bert-large-portuguese-cased-assin-similarity	0.859	0.3
ruanchaves/mdeberta-v3-base-assin-similarity	0.855	0.39
ruanchaves/bert-base-portuguese-cased-assin-similarity	0.847	0.33

ASSIN - RTE ( Recognizing Textual Entailment )

Model	Accuracy	F1
ruanchaves/mdeberta-v3-base-assin-entailment	0.927	0.862
ruanchaves/bert-large-portuguese-cased-assin-entailment	0.92	0.828
ruanchaves/bert-base-portuguese-cased-assin-entailment	0.92	0.827

HateBR ( Offensive Language Detection )

Model	Accuracy	F1
ruanchaves/bert-large-portuguese-cased-hatebr	0.928	0.928
ruanchaves/mdeberta-v3-base-hatebr	0.916	0.916
ruanchaves/bert-base-portuguese-cased-hatebr	0.914	0.914

FaQUaD-NLI ( Question Answering )

Model	Accuracy	F1
ruanchaves/bert-large-portuguese-cased-faquad-nli	0.929	0.93
ruanchaves/mdeberta-v3-base-faquad-nli	0.926	0.926
ruanchaves/bert-base-portuguese-cased-faquad-nli	0.92	0.883

PorSimplesSent ( Text Simplification )

Model	Accuracy	F1
ruanchaves/mdeberta-v3-base-porsimplessent	0.96	0.956
ruanchaves/bert-base-portuguese-cased-porsimplessent	0.942	0.937
ruanchaves/bert-large-portuguese-cased-porsimplessent	0.921	0.913

Citation

Our research is ongoing, and we are currently working on describing our experiments in a paper, which will be published soon. In the meanwhile, if you would like to cite our work or models before the publication of the paper, please use the following BibTeX citation for this repository:

@software{Chaves_Rodrigues_eplm_2023,
author = {Chaves Rodrigues, Ruan and Tanti, Marc and Agerri, Rodrigo},
doi = {10.5281/zenodo.7781848},
month = {3},
title = {{Evaluation of Portuguese Language Models}},
url = {https://github.com/ruanchaves/napolab},
version = {1.0.0},
year = {2023}
}