Dataset Overview

The Calendar Scheduling Dataset is a multi-turn dialogue dataset released by NVIDIA in 2025, designed for understanding natural language scheduling constraints and following multi-turn instructions to resolve conflicts.

This dataset consists of multi-turn dialogue data generated through synthetic methods, aimed at training large language models in event scheduling under complex constraints. Based on preset character personas, the data generation process simulates users proposing event requirements—including duration and time slot constraints—in random order; assistants must synthesize all constraint conditions to complete the scheduling task. Released as part of the NVIDIA NeMo Gym framework, this dataset aims to support the construction of reinforcement learning environments with verifiable rewards (RLVR), helping the community enhance model capabilities in adhering to multi-turn instructions.

Dataset Composition

The dataset primarily includes the following fields and structures:

• conversation: Multi-turn dialogue records containing interactions between the user and assistant, ending with the user's final message.
• exp_cal_state: A dictionary format field listing 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: train and validation, totaling 4,000 samples. The specific distribution is as follows:

• Train set (train): 3,872 samples
• Validation set (validation): 128 samples

Data storage uses plain text format, with an approximate total size of 43 MB.