Nemotron Terminal Synthetic Tasks is a dataset released by NVIDIA in 2026, focused on evaluating and training the skill performance of autonomous agents in real Linux terminal environments. It includes tasks based on skill synthesis, aiming to enhance agents' reasoning and execution capabilities in complex command-line interactions.

Dataset Composition

Each task in the dataset is stored in an independent directory, adopting a structured task organization approach. It mainly consists of the following three components:

• Instruction: The task objective is described in natural language via the instruction.md file, including specific requirements, file paths, and constraints, guiding the agent to complete the designated task.
• Environment Context: Utilizes standardized Docker environments, providing 9 pre-built domain-specific images covering scenarios such as data science and security. Agents execute tasks in isolated container environments and tmux sessions, sharing base images to reduce resource overhead.
• Verification Suite: Checks whether tasks are completed through tests/test.sh scripts and programmatic verification mechanisms such as pytest. Verification includes exit codes, existence of output files, compliance of file formats with requirements, and accuracy of numerical results.