Calendar Scheduling is a multi-turn dialogue dataset released by NVIDIA in 2025, designed to understand natural language scheduling constraints and follow multi-turn instructions to resolve conflicts.

This dataset contains multi-turn dialogue data generated through synthetic methods, used to train large language models in event scheduling capabilities under complex constraints. The data generation process is based on preset character personas, simulating users who propose event requirements with duration and time-slot constraints in random order, while the assistant must integrate all constraints to complete the scheduling. This dataset is released as part of the NVIDIA NeMo Gym framework, aiming to support the construction of training environments for verifiable reward reinforcement learning (RLVR) and help the community improve models' multi-turn instruction-following capabilities.

Dataset Composition

This dataset mainly includes the following fields and structure:

• conversation: Multi-turn dialogue records, containing the interaction process between the user and the assistant, ending with the user's last message.
• exp_cal_state: Dictionary format, containing the list of events in the calendar along with their associated durations and constraints, which can be used to verify the accuracy of scheduling results.

The dataset is divided into two subsets, the training set and the validation set, with a total sample size of 4,000 entries. The specific distribution is as follows:

• Training set (train): 3,872 entries
• Validation set (validation): 128 entries

The data storage format is text, with a total storage size of approximately 43 MB.