This dataset was established in March 2025 and generated using a Python simulation platform. It is primarily designed for human-machine collaborative unmanned swarm air combat control in dynamic unknown environments, aiming to provide data support for multi-agent reinforcement learning, human intention embedding, hierarchical meta-command decision-making, and OOD generalization capability evaluation. The dataset is generated based on the Harfang 3D air combat simulation platform and mainly includes 6v6 training scenarios, as well as symmetric scenarios such as 6v6, 8v8, 12v12, and 20v20, and asymmetric OOD test scenarios such as 6v8, 8v12, and 12v20. The dataset contains a single TXT text file, namely replay_buffer.txt. This file records the reinforcement learning elements—states, actions, rewards, etc.—during the unmanned swarm's mission execution. It comprises four columns of data: the swarm state, individual swarm actions, the team reward function value, the next swarm state, and a flag indicating whether the current task is completed (done). The data content includes UAV states, local observations, continuous control actions, meta-command selections, human preference values, reward feedback, task termination information, as well as evaluation metrics such as win rate, damage ratio, and average episode length. The dataset can support training of human-machine collaborative swarm control policies and cross-scenario generalization performance testing.