This section presents a practical evaluation of multi-brand models trained with multiple tokens fine-tuned on the CodeContests dataset, translating and adapting their description into Spanish and expanding conclusions applicable to the real world. The objective was to measure code comprehension and solution generation capability under different temperature values, and to extract useful insights for custom software projects and custom applications.
Method and test data: the CodeContests dataset was used, containing programming problems with inputs, expected outputs, and test cases. Pre-trained models with multi-token support were fine-tuned while maintaining cross-validation and partitioning by difficulty. A temperature sweep was performed for inference, from conservative to high values, evaluating deterministic generation versus creative generation.
Metrics used: output accuracy, pass at k for different k values, correct execution rate on test cases, CodeBLEU for syntactic and semantic comparison, and inference times to estimate production latency. Common failures were also recorded, such as off-by-one errors, type errors, and inappropriate library usage.
Key results: low temperature values improve accuracy and correct execution rate, favoring reproducible solutions suitable for custom software and critical projects. Medium temperatures increase diversity and allow solving less obvious variants of problems, useful for artificial intelligence prototypes and research. High temperatures produce greater creativity but increase syntactic errors and execution failures. The tradeoff between creativity and robustness is critical when designing AI agents for business environments.
Practical implications: for production deployments in aws and azure cloud services and cybersecurity solutions, it is advisable to prioritize low temperatures and automatic output verification policies. For exploratory generation tasks in business intelligence services or AI prototypes for companies, controlled medium temperatures and human validation pipelines are recommended. In all cases, integration with automated testing systems and tools such as power bi for result visualization improves reliability.
Fine-tuning best practices: ensure dataset quality and balance, use validation with real test cases, measure business impact metrics in addition to technical metrics, and optimize latency to meet user experience requirements. Also consider control and analysis mechanisms to minimize cybersecurity risks and sensitive data filtering when using models in production environments.
Business application and Q2BSTUDIO services: at Q2BSTUDIO we are a software development company specialized in custom software and custom applications. We offer comprehensive solutions that combine artificial intelligence, AI for businesses, and AI agents with aws and azure cloud services, cybersecurity, and business intelligence services. We design fine-tuning pipelines and custom model deployment to solve specific business challenges, integrate automated code generation capabilities, and ensure security and compliance controls. We also develop dashboards with power bi to monitor performance metrics and business results.
Recommended use cases: automation of code component generation, code review assistance, unit test generation, and creation of artificial intelligence solution prototypes integrated with aws and azure cloud services. For companies requiring robust solutions, we offer cybersecurity audits and model optimization for production.
Conclusion: the practical evaluation on CodeContests demonstrates that multi-token fine-tuning improves code generation and comprehension capability when combined with appropriate temperature and validation strategies. Q2BSTUDIO brings expertise to transfer these results to real-world custom software projects, custom applications, applied artificial intelligence, cybersecurity, business intelligence services, and aws and azure cloud deployments, also offering integration with AI agents and power bi to maximize value for the company.
Contact Q2BSTUDIO to design a custom solution that combines fine-tuned models, automated testing, and cloud services, optimizing results and reducing production risks.
