Nemotron-SWE-v1 is a large-scale instruction-tuned dataset for software engineering agents, released by NVIDIA in 2025. The associated paper is available at Training Software Engineering Agents and Verifiers with SWE-Gym, aiming to enhance large language models' capabilities in code understanding, multi-step planning, and repair within real-world software development lifecycles.

The dataset is based on problem descriptions from SWE-Gym and R2E-Gym-Subset. Agent interaction trajectories were synthesized using the Qwen3-Coder-480B-A35B-Instruct model via the OpenHands framework. It is primarily used for Supervised Fine-Tuning (SFT) to help models handle software engineering tasks similar to those in SWE-Bench.

Dataset Composition

The dataset contains the r2e_gym subset, which mainly includes the following fields:

• messages: Records the sequence of messages during agent-environment interactions, including task descriptions, model responses, and tool interaction content.
• license: Records the license information for the corresponding sample.
• used_in: Records which models or projects the sample was used in.
• tools: Records the tools callable by the agent and their definitions.
• dataset: Records the original dataset source of the sample.
• repo: Records the name of the associated code repository.