Skip to main content
A run produces two kinds of files: results, which hold the score and each task’s grade, and outputs, which are the files your agent wrote. Every command is in the valkyrie run reference. To download from Python, see Manage runs with Python. In the commands on this page, replace RUN_ID with the run’s ID, which valkyrie run start prints, and TASK_ID with a task’s ID.

​
Download results

This writes the run’s score and each task’s grade to ./results-RUN_ID.json. Choose another file with --path. To score part of the run, name the tasks. Valkyrie recomputes final_score over only those tasks, and the stored result stays as it was:
--s3 saves the results to the run’s storage instead of downloading them. Combined with --task-ids, it replaces the run’s stored result with the subset. To restore the full result, run it again without --task-ids.

​
Download agent outputs

This downloads and extracts each task’s final_output and output artifacts into a new directory named after the benchmark, the agent, and the run, such as ./swebench_my-agent_RUN_ID. Choose the directory with --output-dir, or only some tasks with --task-ids:

​
Download one file or folder

List the paths stored for a run, one page at a time. Pass the returned next_cursor to --cursor for the next page:
Then download one task’s folder, or one file by its full path:
The destination must be a new directory, and defaults to ./RUN_ID. Paths keep their place relative to the run inside it.