At Maritaca, research is an essential part of our DNA. We continuously invest in developing new technologies, focusing on natural language processing (NLP), Portuguese language models and responsible AI.
Our team actively contributes to the scientific community with papers, studies and collaborations that push the boundaries of what artificial intelligence can do — always with attention to local context and the cultural and linguistic specifics of Brazil.
This technical report introduces Sabiá-4 and Sabiazinho-4, a new generation of language models focused on Brazilian Portuguese. The models were developed in four stages: continued pre-training on Portuguese and Brazilian legal corpora, long-context extension to 128K tokens, supervised fine-tuning, and preference alignment.
Benchmark for evaluating models' ability to find information on the web via breadth-first search, with questions paired with checklists to assess completeness and correctness.
Brazilian literature benchmark for evaluating deep comprehension of canonical works by Portuguese language models.
Evaluation set for Brazilian Portuguese text-generation quality, covering creative, technical and journalistic writing.
Benchmark for measuring ideological/opinion bias and sycophancy (excessive agreement) in language models, with comparative analysis of commercial and open models.
Dataset evaluating LLM comprehension of Brazilian proverbs, addressing gaps in existing benchmarks for linguistic and cultural nuances of Portuguese.
Scalable methods to generate high-quality Portuguese corpora from Common Crawl. Result: 120 billion tokens.
Automatic evaluation of legal writing using questions from the 2nd phase of Brazil's Bar exam.
Benchmark with over 17,000 questions about global events to evaluate temporal knowledge of models.
Investigation of pre-training laws for models of varying sizes on general vs. specialized data.
Technical report for Sabiá-3, the previous generation of the Sabiá family.
Sabiá-2 models with advances in fluency and benchmark performance on Brazilian tasks.
First technical report of the Sabiá family, showing significant gains in PT-BR over generalist models.
First LLM trained on Brazilian legal data from reputable sources.
Study showing Chain-of-Thought gains on Brazilian college-entrance exams.
Multimodal benchmark based on USP and UNICAMP entrance exams.
Maritaca believes that the advancement of artificial intelligence must walk hand in hand with education and open research. That's why we offer free API credits to students, professors and researchers who want to use the Sabiá family — LLMs specialized in Portuguese — in teaching or scientific projects.
This initiative seeks to democratize access to our technology, contributing to the training of new professionals, the development of relevant solutions and the production of knowledge in the Portuguese AI ecosystem.
