Nemotron-Cascade-SFT-SWE

## Dataset Description: The Nemotron-Cascade-SFT-SWE dataset is the RL training data for SWE code repairing task, consisting of [SWE-Bench-Train](https://huggingface.co/datasets/princeton-nlp/SWE-bench), [SWE-reBench](https://huggingface.co/datasets/nebius/SWE-rebench), [SWE-Smith](https://huggingface.co/datasets/SWE-bench/SWE-smith), [R2E-Gym/R2E-Gym-Subset](https://huggingface.co/datasets/R2E-Gym/R2E-Gym-Subset) and [SWE-Fixer-Train](https://huggingface.co/datasets/internlm/SWE-Fixer-Train-110K). We select the training data for SFT and RL stages based on its difficulty. Also, to avoid data contamination, we exclude all instances originating from repositories present in the [SWE-Bench_Verified](https://huggingface.co/datasets/princeton-nlp/SWE-bench_Verified) evaluation dataset. We create the prompts following the [agentless mini framework](https://github.com/facebookresearch/swe-rl), consisting of three tasks buggy code localization, code reparing and test case generation. The response is generated using [DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) Using the data for [SFT](https://huggingface.co/datasets/nvidia/Nemotron-Cascade-SFT-SWE) and [RL](https://huggingface.co/datasets/nvidia/Nemotron-Cascade-RL-SWE), we reach pass@1 resolve rate (without TTS) **37.2** and **43.1** for 8B and 14B models. See more detail of data creation in our [technical report](https://arxiv.org/abs/2512.13607). Note that All SWE datasets are upsampled by 3× before being incorporated into the [Stage-2 SFT data blend](https://huggingface.co/datasets/nvidia/Nemotron-Cascade-SFT-Stage-2). The following are the statistics for our training data: **Code Localization** | **Source** | **# Question** | **# Sample** | | :---: | :---: | :---: | | [SWE-Fixer-Train](https://huggingface.co/datasets/internlm/SWE-Fixer-Train-110K) | 52,702 | 53,230 | | [SWE-Smith](https://huggingface.co/datasets/SWE-bench/SWE-smith) | 9,707 | 9,714 | | [SWE-Bench-Train](https://huggingface.co/datasets/princeton-nlp/SWE-bench) | 10,251 | 16,184 | | [SWE-reBench](https://huggingface.co/datasets/nebius/SWE-rebench) | 9,654 | 9,693 | | [R2E-Gym/R2E-Gym-Subset](https://huggingface.co/datasets/R2E-Gym/R2E-Gym-Subset) | 3,444 | 3,444 | **Code Repairing** | **Source** | **# Question** | **# Sample** | | :---: | :---: | :---: | | [SWE-Fixer-Train](https://huggingface.co/datasets/internlm/SWE-Fixer-Train-110K) | 22,311 | 33,301 | | [SWE-Smith](https://huggingface.co/datasets/SWE-bench/SWE-smith) | 2,822 | 17,625 | | [SWE-Bench-Train](https://huggingface.co/datasets/princeton-nlp/SWE-bench) | 3,660 | 19,540 | | [SWE-reBench](https://huggingface.co/datasets/nebius/SWE-rebench) | 2,822 | 16,441 | **Test Case Generation** | **Source** | **# Question** | **# Sample** | | :---: | :---: | :---: | | [SWE-Fixer-Train](https://huggingface.co/datasets/internlm/SWE-Fixer-Train-110K) | 18,792 | 19,678 | | [SWE-Smith](https://huggingface.co/datasets/SWE-bench/SWE-smith) | 2,731 | 2,881 | | [SWE-Bench-Train](https://huggingface.co/datasets/princeton-nlp/SWE-bench) | 4,797 | 6,066 | | [SWE-reBench](https://huggingface.co/datasets/nebius/SWE-rebench) | 3,021 | 3,026 | **Dataset Structure** An example of datum is as follows: ``` category: (str) - Task; e.g., SWE Repair / SWE Localization / SWE TestGen . source: (str) - The dataset name where the instance originally comes from. messages: (list) - The user prompt and model response for SWE Bench following the agentless mini framework. generator: (str) - The model we used for response generation patch: (str) - Comments made on the issue prior to the creation of the solution PR’s first commit creation date. thinking: (bool) - Thinking or non thinking mode. ```