BRAT: An Intent-to-Kubernetes Translation via LLM Fine-Tuning

The widespread adoption of cloud computing technologies enabled notable consolidation of computational resources, with Kubernetes (K8s) as the default open-source container orchestration platform that automates application management. Since properly configuring K8s requires writing manifest files, i.e., YAML-structured files that define the containerized environment, the advent of Large Language Models (LLMs) can automate this process by relaxing user input to an intent expressed in natural language. However, fine-tuning an LLM poses unique challenges and reveals gaps in domainspecific knowledge, as the generated output must be both semantically and syntactically correct, and the area lacks a high-quality dataset. To address these challenges, we propose BRAT, a pipeline that generates K8s manifests from user-defined intents using a fine-tuned LLM. First, we created a dataset of intent-K8s pairs, leveraging advanced LLM and applying n-shot learning. The adoption of those models revealed that different models require personalized prompting, increasing costs by up to 66%. Then, leveraging the newly created dataset, we fine-tuned Granite-3B, chosen for its lightweight capabilities, to generate accurate manifest files, resulting in an average improvement of 110% over the baseline.