by NVIDIA
Open-SWE-Traces is an NVIDIA dataset of agent trajectories for fixing real software issues, built for fine-tuning and distilling software engineering agents.
DatasetFree
Step 1: Open the dataset page on Hugging Face and read the data fields table.
Step 2: Load the dataset with the Hugging Face datasets library.
Step 3: Filter trajectories by the resolved field, framework or repository language.
Step 4: Use the messages and tools fields as training examples for supervised fine-tuning or distillation.
Step 5: Cite the paper listed on the dataset card if you publish results.
CC BY 4.0. The card says users must check the content and applicable licenses for their own use.
The complete conversation history with system, user, assistant and tool messages, the tool definitions available to the agent, and a flag showing whether the task was resolved.
OpenHands, SWE-agent, mini-swe-agent, opencode and claude-code.
Issue statements and repository environments are taken from permissively licensed datasets, Scale-SWE and SWE-rebench-V2.