finding-lobbench-results

To avoid wasting GPU hours on re-benching, you should search existing pipeline directories first. This Skill checks agent_outputs, kangli lob_pipeline, and aramis lob_pipeline to locate prior results for your checkpoint or experiment.

Yes, you can locate LOBbench results using a SLURM job ID or wandb run. The Skill accepts these identifiers to apply a strict search sequence across pipeline directories and retrieve the relevant horizon and metrics from matching output files.

The LOBbench search sequence checks agent_outputs, kangli lob_pipeline, and aramis lob_pipeline directories. It scans these specific locations to identify matching results for your provided checkpoint, SLURM job, or experiment identifier.

The LOBbench results report includes the exact file location of any matching results alongside the relevant horizon and metrics extracted from agent outputs or summary files. If no prior results exist, it indicates that no results were found.

Checking for existing LOBbench results prevents unnecessary computations and saves GPU hours. Prior bench runs for the same checkpoint or experiment are often already available in pipeline directories and can be reused instead of re-running the bench.