A Dataset of LLM Strategic Play

This dataset aims to support the systematic evaluation of Large Language Models (LLMs) in structured decision-making environments. It contains 9,572 move-level records generated from competitive interactions among twenty LLMs across four board games—Tic-Tac-Toe, Connect Four, Suicide, and Not Connect Four—collected during an eight-round Swiss-System tournament conducted between May and July 2025. The data captures the full context of each decision, including board states encoded in GDL, legal move sets, model identifiers, execution times, and the raw reasoning provided by each agent. A GDL-based validation pipeline, including a multi-step correction mechanism, ensures the legality and integrity of all recorded actions. The scope of the dataset is to enable analyses of strategic behavior, rule adherence, performance stability, and model-to-model variability, providing a resource for benchmarking and studying LLM behavior in controlled, reproducible game environments.