Raw game records and multi-level instruction datasets populate this repository, focusing on the strategic interactions of the Werewolf game. The data includes event.json files that pair chronological game records with the internal thinking processes of the agents.
Use Cases
- Train models to simulate player logic by utilizing the thinking process data within the event.json files.
- Develop multi-agent reinforcement learning agents using the game regular records to model strategic interactions.
- Fine-tune Large Language Models for social deduction tasks using the multi-level instruction datasets.
Strengths
- Includes event.json files that store both game regular records and internal thinking process data.
- Features multi-level instruction datasets processed for reinforcing strategic interactions in Large Language Models.
- Provides raw game data used in the 'Multi-agent KTO' research paper for reinforcing agent interactions.