Manacá-1B: An Open, Reproducible Brazilian-Portuguese Language Model and a Tokenizer-Aware, Paired Evaluation
为了解决巴西葡萄牙语开放语言模型缺乏和难以复现的问题，本文通过完全容器化的可复现流程训练并发布了Manacá-1B模型，并对其进行了全面评估。
0 citationsRead paper
Selected work
Representative Papers
A Novel Procedural Generation for Level Design of Mansions and Dungeons
This work proposes a three-stage indoor scene generation method that integrates established level design principles to address the spatial disorganization and poor playability often found in traditional procedural content generation. The approach begins with binary space partitioning (BSP) to construct an initial layout, followed by graph traversal algorithms to ensure logical room connectivity. A final post-processing stage enhances structural coherence and visual consistency. By systematically embedding design principles into each phase of the pipeline, the method preserves architectural plausibility while improving navigability and diversity. Experimental results demonstrate high flexibility and efficacy: across 100,000 generated maps, over 91% achieved full connectivity when appropriate parameters were used, confirming the robustness of the proposed framework.
0 citationsRead paper
Recent publications
Latest Papers
Manacá-1B: An Open, Reproducible Brazilian-Portuguese Language Model and a Tokenizer-Aware, Paired Evaluation
为了解决巴西葡萄牙语开放语言模型缺乏和难以复现的问题，本文通过完全容器化的可复现流程训练并发布了Manacá-1B模型，并对其进行了全面评估。
0 citationsRead paper
A Novel Procedural Generation for Level Design of Mansions and Dungeons
This work proposes a three-stage indoor scene generation method that integrates established level design principles to address the spatial disorganization and poor playability often found in traditional procedural content generation. The approach begins with binary space partitioning (BSP) to construct an initial layout, followed by graph traversal algorithms to ensure logical room connectivity. A final post-processing stage enhances structural coherence and visual consistency. By systematically embedding design principles into each phase of the pipeline, the method preserves architectural plausibility while improving navigability and diversity. Experimental results demonstrate high flexibility and efficacy: across 100,000 generated maps, over 91% achieved full connectivity when appropriate parameters were used, confirming the robustness of the proposed framework.
0 citationsRead paper