This dataset is for evaluating the task of Black-box Multi-agent Integration which focuses on combining the capabilities of multiple black-box conversational agents at scale. It provides data to explore two main frameworks of exploration: question agent pairing and question response pairing. Overall this dataset contains 5550 utterances with 19 question-response pairs per question (one from each of the 19 agents), 105,450 in total across 37 domains. The utterances are split into 3700 utterances (100 examples per domain) for the training set and 1850 (50 per domain) for the test set. The train and test sets respectively contain 2399 and 1186 utterances with at least one positive question-response pair. In the remaining examples, none of the agents were able to achieve annotator agreement (>= 3).
To evaluate our proposed strategy of asynchronous communication for LLMs, we run games of Mafia with human players, incorporating an LLM-based agent as an additional player, within an asynchronous chat environment. The data includes all players messages and votes including timestamps, game management messages (e.g., announcements of the beginning and end of phases), in addition to records related to the agent, such as the prompts that were provided at each timestamp. While prior datasets for Mafia exist, ours is the first to integrate an LLM agent, allowing analysis of human-LLM interactions. Detailed analysis of the data is available in our paper: Time to Talk: LLM Agents for Asynchronous Group Communication in Mafia Games