Large Language Models (LLMs) play a vital role in contemporary software development processes, particularly for code completion, generation, and debugging tasks. These AI systems support programmers by delivering precise, context-sensitive recommendations that optimize coding workflows and substantially increase development efficiency. This research concentrates on examining the Stack-v2 dataset to detect complex code patterns that typically result in inadequate completions from the model. We tackle these issues by refining the StarCoder-15B model with a carefully chosen portion of Stack-v2, specifically targeting challenging prediction scenarios. After fine-tuning, we assessed the model using the JavaScript component of the HumanEval-X benchmark. Our findings indicate enhancements compared to the original model, reaching a pass@1 score of 34.4 points. This confirms the efficacy of our focused fine-tuning method for improving LLMs to deliver more dependable and sturdy code completion in practical applications.