This dataset was constructed in June 2025 using a Python simulation platform and the MPE (Multi-Agent Particle Environment) navigation task framework, with automated data collection facilitated by Wandb. It is designed to compare the performance differences between two credit assignment mechanisms—Shapley Dividend and the conventional Shapley Value—when embedded into multi-agent reinforcement learning algorithms (SD3PG and SQDDPG). All experiments were conducted under identical network architectures and hyperparameter configurations. For each set of conditions, data were collected across five distinct random seeds, recording the per-episode average reward and collision count during training. Summary statistics, including maximum, minimum, and mean values, were computed to comprehensively reflect the expected performance, optimal bounds, and robustness of the algorithms. The dataset comprises two CSV table files, namely mean_end_reward.csv and mean_ep_collisions.csv, along with additional visualization and derivative metric files, such as training reward curves, collision trend plots, and convergence speed analysis tables. By comparing convergence speed, peak final rewards, and fluctuation ranges across training curves, it can be verified that Shapley Dividend, relative to Shapley Value, more effectively guides agents to form cooperative strategies in navigation tasks. It not only accelerates the evolution of cooperative behavior and reduces collision frequency, but also achieves superior and more stable final collaborative performance. This dataset provides reusable benchmark data and analytical foundations for research on credit assignment mechanisms in multi-agent systems.