Polygl0t/Tucano2-qwen-3.7B-Think

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 combination of supervised fine-tuning (SFT) and Anchored Preference Optimization (APO). A key differentiator of this model is its specialization in reasoning, designed to generate Chain-of-Thought (CoT) traces within <think> and </think> special tokens, providing insight into its thought process.

Key Capabilities

• Portuguese Language Modeling: Primarily designed for interaction and research in the Portuguese language.
• Reasoning Focus: Fine-tuned to produce CoT-style reasoning traces, enhancing interpretability for complex tasks.
• Open and Reproducible: All datasets, source code, and training recipes for the Tucano2 series are publicly available.
• Performance: Achieves a Normalized Performance Metric (NPM) of 54.07 in Knowledge & Reasoning benchmarks, outperforming SmolLM3-3B and Qwen3-4B in this category.

Intended Uses

• Research and Development: Serves as a foundation for advanced Portuguese language modeling research.
• Fine-tuning Base: Can be adapted and fine-tuned for specific real-world applications under the Apache 2.0 license.

Limitations

• Hallucinations and Biases: Inherits common LLM limitations such as generating false information and exhibiting social biases.
• Language Specificity: Primarily optimized for Portuguese; performance in other languages may be limited.
• Repetition: May exhibit repetition loops or verbosity, especially with suboptimal generation parameters.
• No Coding Focus: Not trained on coding data, therefore not recommended for code generation tasks.