Nemotron-SWE-v2 is an instruction-tuned dataset for the software engineering (SWE) domain, released by NVIDIA in 2025. The related paper is available at "Training Software Engineering Agents and Verifiers with SWE-Gym", and the dataset aims to enhance the capabilities of large language models in code repair, localization, and test generation for SWE-Bench-style tasks.

The dataset contains a total of 256,254 samples, covering agentless software engineering task data and agent-based multi-step interaction trajectories. It is suitable for supervised fine-tuning (SFT) and model distillation, and can be used to train software engineering assistants with capabilities in code analysis, tool calling, and autonomous problem-solving. The dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0), with supplementary licenses including Apache 2.0, MIT, BSD-3-Clause, and BSD-2-Clause.

Dataset Composition

• Agentless SWE: Contains 209,976 samples. This is supervised fine-tuning data for agentless software engineering subtasks, covering code localization, code repair, and test generation. The data was generated by DeepSeek-R1-0528, with each example including task-specific outputs such as ranked file lists, patch edits, and unit tests.
• Agentic SWE: Contains 46,278 samples. This subset includes agent trajectories collected using the OpenHands framework and synthesized by Qwen3-Coder-480B-A35B-Instruct, aiming to improve model performance through multi-step tool usage and problem-solving. Problem statements are sourced from SWE-Gym and R2E-Gym-Subset.