Аннотация
The emergence of advanced open-source large language models (LLMs) opens new opportunities for researchers and accelerates the development of generative AI. This study focuses on harnessing the potential of Gemma, Google's recently released open source LLM, to develop the Kaggle assistant through fine-tuning. The main objective of the study was to investigate the Low-Rank Adaptation (LoRA) fine-tuning method of Gemma model for building a robust AI assistant in Kaggle. Using the Gemma 2B variant, this study applies the Latent Consistency Model - LoRA method to improve interpretability and performance. The fine-tuning of the model achieved a sparse categorical cross-entropy (Loss) of 0.4943 and a sparse categorical precision of 0.7753. Although experiments with epochs and batch sizes revealed potential overbuilding after 10 epochs, the model shows acceptable performance in this range. The integration of LoRA fine-tuning enhances the model's ability to provide accurate recommendations and conclusions to users solving Kaggle problems, thereby improving the user experience and reinforcing its utility as a valuable resource for data scientists. Further refinement and adaptation of the Gemma model, especially through techniques such as LoRA fine-tuning, are critical to ensure its continued effectiveness in assisting Kaggle users in their data science endeavors.