Atlantica 1B PT-BR Instruct Overview

Atlantica-1B-PT-BR is a 1 billion parameter instruction-tuned language model developed by thejpdev, specifically fine-tuned for Brazilian Portuguese. It is built upon the unsloth/llama-3.2-1b-instruct-bnb-4bit base model, with its behavior significantly adapted through supervised fine-tuning using a custom dataset of approximately 572,000 samples.

Key Capabilities

• Instruction Following: Designed to accurately follow instructions provided in Brazilian Portuguese.
• Question Answering: Capable of responding to open-ended and contextual questions.
• Text Classification: Performs simple text classification tasks.
• Educational Tasks: Suitable for educational and benchmark-style tasks in Portuguese.
• Local Inference: Optimized for lightweight local inference, compatible with tools like LM Studio using GGUF versions.

Evaluation and Use Cases

The model has been evaluated using Portuguese benchmarks such as ENEM, TweetSentBR, and Bluex, showing consistent performance for a small model trained for Portuguese instructions. Due to its compact size, Atlantica-1B-PT-BR is particularly well-suited for general-purpose applications and environments with limited computational resources.