Model Card for GAIA (Gemma-3-Gaia-PT-BR-4b-it)

GAIA is an open, state-of-the-art language model for Brazilian Portuguese. It was developed by continuously pre-training the google/gemma-3-4b-pt model on an extensive, high-quality corpus of Portuguese data.

The goal of GAIA is to democratize access to cutting-edge AI technology in Brazil, enabling developers, researchers, and organizations to build innovative solutions on a robust and reliable technological foundation.

Model Details

Model Description

GAIA was developed through a partnership between The Brazilian Association of AI (ABRIA), the Center of Excellence in Artificial Intelligence (CEIA) at the Federal University of Goiás (UFG), startups Nama and Amadeus AI, and Google DeepMind.

The development process started with the base model google/gemma-3-4b-pt and involved two main stages:

1. Continuous Pre-training: The model was trained on a large, high-quality Portuguese dataset totaling approximately 13 billion tokens. This corpus includes a variety of domains, such as scientific articles and Wikipedia data in Portuguese, ensuring a deep understanding of the language and its contexts.
2. Instruction-Following Capability Restoration: To enable the model to follow instructions without traditional supervised fine-tuning (SFT), a weight merging operation was applied. This technique, described in the paper “Balancing Continuous Pre-Training and Instruction Fine-Tuning: Optimizing Instruction-Following in LLMs”, allows the model to integrate the knowledge acquired during continuous pre-training with the ability to interact in a chat format and follow instructions.

• Developed by: The Brazilian Association of AI (ABRIA), the Center of Excellence in Artificial Intelligence (CEIA-UFG), Nama, Amadeus AI, and Google DeepMind.
• Model: GAIA
• Model type: Causal decoder-only Transformer-based language model.
• Language(s): Brazilian Portuguese (pt-BR)
• License: Gemma
• Based on: google/gemma-3-4b-pt

Team

This project was made possible by the contributions of the following individuals:

• Dr. Celso Gonçalves Camilo-Junior
• Dr. Sávio Salvarino Teles de Oliveira
• Me. Lucas Araujo Pereira
• Marcellus Amadeus
• Daniel Fazzioni
• Artur Matos Andrade Novais
• Salatiel Abraão Avelar Jordão

Model Sources

• Repository: CEIA-UFG/Gemma-3-Gaia-PT-BR-4b-it
• Paper (Merge Methodology): Balancing Continuous Pre-Training and Instruction Fine-Tuning: Optimizing Instruction-Following in LLMs

Uses

The model is designed for text generation and conversational tasks in Portuguese.

Direct Use

GAIA can be used directly for chat, question answering, summarization, creative content generation, and other tasks requiring natural language understanding and generation in Portuguese.

Downstream Use

GAIA serves as an excellent base model for fine-tuning on specific tasks, such as:

• Sentiment analysis in Portuguese.
• Retrieval-Augmented Generation (RAG) systems for corporate knowledge bases.
• Document classification.
• Specialized customer service chatbots.

Out-of-Scope Use

This model should not be used for high-stakes, critical decisions without human oversight. Its use for generating malicious, offensive, or illegal content, or for deceptively impersonating a human, is outside the intended scope. The model's performance in languages other than Portuguese will be significantly degraded.

Bias, Risks, and Limitations

Like any language model, GAIA reflects the biases present in its training data. Although the training corpus was curated with a focus on high quality, it may contain social and cultural biases from sources like Wikipedia and scientific articles. Therefore, the model may generate content that perpetuates existing stereotypes.

Furthermore, the model can "hallucinate," meaning it can generate information that appears factual but is not true. We strongly recommend verifying critical facts generated by the model before any use.

Recommendations

Users (both direct and downstream) should be aware of the model's risks, biases, and limitations. Implementing safeguards and content moderation is recommended, especially in public-facing applications. Human supervision is crucial for sensitive use cases.

Training Details

Training Data

The continuous pre-training was performed on a corpus of approximately 13 billion tokens in Portuguese. The data selection prioritized high quality and diversity, including sources such as:

• Scientific Articles in Portuguese: To provide the model with more formal and technical knowledge.
• Portuguese Wikipedia: To cover a wide range of general knowledge.

A rigorous cleaning and filtering process was applied to ensure the highest possible data quality.

Training Procedure

The training was conducted on a DGX infrastructure with NVIDIA H100 GPUs, using between 3 and 5 GPUs in parallel.

Training Hyperparameters

• Training regime: Mixed Precision (bf16)
• Global Batch Size: 4 million tokens

Evaluation

The model was evaluated on a set of multiple-choice benchmarks in Portuguese, comparing its performance against the base model, google/gemma-3-4b-it. The benchmarks include BlueX (a compilation of multiple-choice questions), and questions from the ENEM (Brazilian High School National Exam) and OAB (Brazilian Bar Exam).

Results

Benchmark	google/gemma-3-4b-it (Baseline)	GAIA (Our Model)
BlueX	0.6630	0.6575
ENEM 2024	0.6556	0.7000
ENEM (General)	0.7416	0.7486I
OAB (Bar Exam)	0.4502	0.4416

Summary

The results indicate that continuous pre-training on Portuguese data had a notable impact on the model's performance. GAIA showed a significant improvement on the ENEM 2024 benchmark, outperforming the Google base model. On other benchmarks like BlueX and OAB, its performance is competitive and very close to the original model's, suggesting that the additional training process maintained the model's general capabilities while enhancing its knowledge in specific Portuguese-language domains.

Citation

If you use this model in your research or application, please cite our work.

BibTeX:

@misc{gaia-gemma-3-4b-2025,
    title={GAIA: An Open Language Model for Brazilian Portuguese},
    author={CAMILO-JUNIOR, C. G.; OLIVEIRA, S. S. T.; PEREIRA, L. A.; AMADEUS, M.; FAZZIONI, D.; NOVAIS, A. M. A.; JORDÃO, S. A. A.},
    year={2025},
    publisher={Hugging Face},
    journal={Hugging Face repository},
    howpublished={\url{[https://huggingface.co/CEIA-UFG/Gemma-3-Gaia-PT-BR-4b-it](https://huggingface.co/CEIA-UFG/Gemma-3-Gaia-PT-BR-4b-it)}}
}