Nemotron-Cascade-SFT-SWE is a training dataset for software engineering tasks released by NVIDIA, with the associated paper "Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models". It aims to enhance the capabilities of large language models in software engineering tasks such as code localization, code repair, and test case generation.

The dataset integrates data from multiple sources, including SWE-Bench-Train, SWE-reBench, SWE-Smith, R2E-Gym/R2E-Gym-Subset, and SWE-Fixer-Train, and applies difficulty filtering and sample adjustments based on the training requirements of supervised fine-tuning (SFT) and reinforcement learning (RL). To reduce the risk of evaluation data contamination, the dataset excludes instances originating from code repositories covered by the SWE-Bench_Verified evaluation set.

Dataset Composition

The dataset includes the following three JSONL files:

• swe_localization.jsonl: Code localization task, used to locate code files that need to be inspected or modified.
• swe_repair.jsonl: Code repair task, used to generate code repair solutions.
• swe_testgen.jsonl: Test case generation task, used to generate test cases that verify code behavior.

Each sample mainly contains the following fields:

• category: Task category, including SWE Localization, SWE Repair, and SWE TestGen.
• source: Original data source.
• messages: A list of messages containing user prompts and model responses.
• generator: The name of the model used to generate the model response.
• patch: Comment information related to the relevant Issue, referring to comments made before the first commit of the solution PR.
• thinking: A boolean value indicating whether thinking mode is enabled.