Dataset Introduction

Nemotron-SFT-SWE-v2 is a software engineering instruction fine-tuning dataset designed to improve the capabilities of large language models on SWE-Bench style tasks. This dataset contains agent trajectories collected using the OpenHands framework, as well as agentless SWE subsets for subtasks such as code localization, code repair, and test generation. This dataset is suitable for commercial use, primarily targeting LLM engineers and research teams building autonomous software engineering agents and code assistants. It is suitable for supervised fine-tuning and model distillation, helping models interpret real-world problem statements, plan multi-step tool usage, navigate codebases, and implement fixes in a SWE-Bench style environment.

Dataset composition

subset

The dataset contains two subsets: – SWE trajectory of the intelligent agentApproximately 46k agent trajectories were collected using the OpenHands framework. These trajectories were synthesized using Qwen3-Coder-480B-A35B-Instruct and carefully curated for supervised fine-tuning to improve model performance on the SWE-Bench task. The problem statement is derived from SWE-Gym and R2E-Gym-Subset. – SWE without agentSupervised fine-tuning data for agentless software engineering subtasks in the SWE-Bench style, including code location, code fixing, and test generation. Each example includes task-specific outputs such as a ranked list of located files, fixed patches, and test generation reproducibility and unit tests.

Data Quantification

Data Format

Modality: Text Format: JSONL Structure: Text + Metadata

license

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license. Other licenses include Apache 2.0, MIT, BSD-3-Clause, and BSD-2-Clause.