ubicloud/

filtered-SWE-Openhands-Devstral-32k-20K

This dataset contains 20,000 high-quality SFT trajectories filtered and selected from SWE-Star, a 250K+ sample dataset collected under a modified OpenHands scaffold. The data is used to train SWE agent models under limited resources with a 32K context window. A single-source dataset is intentionally selected so that all training trajectories share a uniform scaffold, isolating the scaffold inconsistency effect: any performance variation across evaluation… See the full description on the dataset page: https://huggingface.co/datasets/ubicloud/filtered-SWE-Openhands-Devstral-32k-20K.