MobaQA: MOBA Games Prediction Based on Large Language Model Fine-Tuning

Multiplayer Online Battle Arena (MOBA) games, one of the most popular esports genres, have been extensively studied in the fields of game AI and game analytics. Open-source large language models (LLMs), such as Llama, have become essential tools in artificial intelligence (AI) research. However, previous battlefield information prediction studies have largely assumed that all in-game data are readily accessible, using sequential models to analyze samples containing numerous data items collected over a period, which present challenges for users to collect and input. In this work, we introduce MobaQA, a MOBA battlefield information prediction model based on LLMs fine-tuning. Instead of utilizing extensive data inputs, MobaQA employs a minimal set of key data items to construct its dataset. By fine-tuning LLMs, MobaQA generates predictions based on input battlefield information text. Using samples collected from two major game versions over a span of two years, we design multiple Q&A templates and conduct extensive experiments on fine-tuned MobaQA models derived from three different base LLMs, under varying training volumes and with various levels of missing data. Experimental results indicate MobaQA can effectively predict real-time winrate and match outcome.