Polygl0t/Tucano2-qwen-3.7B-Think
Polygl0t/Tucano2-qwen-3.7B-Think is a 3.76 billion parameter instruction-tuned Portuguese language model developed by Polygl0t, built on the Qwen3 Transformer architecture. Fine-tuned with supervised fine-tuning (SFT) and Anchored Preference Optimization (APO), it specializes in generating Chain-of-Thought (CoT) reasoning traces within and tokens. This model is optimized for research and development in Portuguese language modeling, particularly for tasks requiring explicit reasoning.
Loading preview...
Tucano2-qwen-3.7B-Think Overview
Polygl0t/Tucano2-qwen-3.7B-Think is a 3.76 billion parameter instruction-tuned Portuguese language model based on the Qwen3 Transformer architecture. It was developed by Polygl0t through a two-stage fine-tuning process involving Supervised Fine-Tuning (SFT) and Anchored Preference Optimization (APO), utilizing the Polygl0t/gigaverbo-v2-sft and Polygl0t/gigaverbo-v2-preferences datasets. A key differentiator is its specialization in generating explicit Chain-of-Thought (CoT) reasoning traces, encapsulated within <think> and </think> special tokens, making its thought process transparent.
Key Capabilities
- Portuguese Language Proficiency: Primarily designed for interaction and understanding of the Portuguese language.
- Reasoning Generation: Excels at producing step-by-step reasoning (CoT) for complex problems, enhancing interpretability.
- Open and Reproducible: All datasets, source code, and training recipes are publicly available for full reproducibility.
- Competitive Performance: Achieves a Normalized Performance Metric (NPM) of 54.07 in Knowledge & Reasoning tasks, outperforming SmolLM3-3B and Qwen3-4B in this category.
Intended Use Cases
- Research and Development: Serves as a foundational model for advanced research in Portuguese NLP.
- Fine-tuning: Suitable for further adaptation and fine-tuning for specific real-world applications under the Apache 2.0 license.
- Educational Tools: Can be used in applications requiring transparent problem-solving steps, such as educational platforms for math or science.
Limitations
Users should be aware of potential limitations including hallucinations, biases inherited from training data, and reduced performance in languages other than Portuguese. It is not optimized for coding tasks; for those, the Instruct variants are recommended.