Open-Source NLP Research Projects

Open-source leaderboard to evaluate LLM memorization, reasoning and linguistic capabilities in the languages of Spain, LATAM and the Caribbean. Developed as part of the #Somos600M Project thanks to the donation of 60 high-quality datasets by 10 research groups. La Leaderboard is live!

The INCLUDE project aims to create a multicultural and multilingual benchmark. As part of my Master's thesis, I'm leading the effort to extend the v1 to the languages of LATAM, the Caribbean and Spain. We're looking for exams in EVERY language and welcome active participation. Share an exam in your language!

FineWeb 2 is an annotation campaign by Hugging Face aimed at creating a high-quality multilingual dataset. I'm the Language Lead for Spanish and responsible for the other languages from LATAM, the Caribbean and Spain. The annotation consists of evaluating the educative level of texts, it's really easy and makes a huge impact. Join the effort!

Community effort to validate the machine-translated Spanish versions of 3 widely-used evaluation datasets (MMLU, RAC-C, and HellaSwag) and the prompt dataset from the Data Is Better Together (DIBT) initiative. Efforts co-organized by SomosNLP, Hugging Face & Argilla. Join us!

At SomosNLP we are collecting datasets in the languages spoken in LATAM, the Caribbean and Spain. Collaborate and help us collect diverse data!

Discover the initiatives driving NLP advancements in Spanish and other low-resource languages spoken in LatAm and Spain.

Third edition of the largest open-source hackathon of NLP in Spanish. This year's edition counted with +600 participants and 12 amazing speakers.
Check the recorded talks and keynotes!

This tool allows you to self-assess the transparency of your model development based on the Foundation Model Transparency Index (FMTI) published by the Center for Research on Foundation Models.

Second edition of the largest open-source hackathon of NLP in Spanish. This year's edition counted with +500 participants, 17 speakers, and 7 mentors.
Check the awarded projects and the recorded talks and keynotes!

Organized two initiatives to promote both the work and research of women in NLP and also the projects that apply NLP to fight sexism.

Contributing to the translation of the NLP Course by Hugging Face to Spanish.

Contributing to BigCode. Project in progress.

Contributing to EleutherAI's research project "Polyglot Romance". Project in progress.

With more than 500 participants from 39 countries, it is the largest open-source hackathon of NLP in Spanish. The recorded talks and workshops have already more than 5k visualizations! Organized by SomosNLP and sponsored by Hugging Face, Platzi and Paperspace. Check the awarded projects!

A one-year long international research workshop on large multilingual models and datasets. We created, among other cool things, ROOTS: A 1.6TB Composite Multilingual Dataset that was then used to train BLOOM: A 176B-Parameter Open-Access Multilingual Language Model.

BERTIN is a series of RoBERTa-based models in Spanish trained using a novel sampling technique that we call "perplexity sampling". More detailed info can be found in the model card and the paper BERTIN: Efficient Pre-Training of a Spanish Language Model using Perplexity Sampling.

The first NLP course from zero to hero in Spanish. It's open-source and was organized by SomosNLP with the support of Spain AI. I taught the classes on sequential models and the Transformer architecture.

A series of Spanish language models trained with Flax/Jax and using TPUs sponsored by Google during the Flax/Jax Community Week organized by Hugging Face in June 2021. Here are the model cards: GPT-2 model , T5 model and Wav2Vec2 model.

A program where we provide women entrepreneurs with the tools, knowledge, mentoring and network to successfully realize their startup/business idea in the AI sector.

Addition of 3 datasets in Spanish to the huggingface/datasets library during the open-sprint organized by Hugging Face in Dec 2020. The datasets are HEAD-QA (a multi-choice HEAlthcare Dataset), the dataset of the eHealth-KD Challenge at IberLEF 2020, and the Spanish Billion Words Corpus.

PyPI package to perform quality analyses on ML models. It focuses on the three quality pillars: functionality, robustness and explainability.

Chatbot that understands and answers questions about the COVID-19: symptoms, prevention, regulation, the situation in Spain. Don't hesitate to chat with AURORA!

The chatbot understands correctly on the 1st attempt 92% of the requests and helped 1500+ people during the first months of the pandemic. Collaboration with Accenture’s Gijón office.

Implementation of a Neural Network that predicts - with a correlation coefficient of 0.778 - characteristics of the Higgs Boson produced in the particle collider. Collaboration with the university's high energy particle research team.