pmlb-batch-inference

To run batch inference across local PMLB datasets, the Skill iterates through pmlb/datasets using your existing single-dataset script, appending dataset-by-dataset metrics to a strict-column CSV file without overwriting previous results.

Yes, you can resume batch inference runs. The workflow reads the dataset column from existing result CSVs to skip completed items, appending new metrics for unprocessed datasets without overwriting previous outputs.

You apply multiplicative Gaussian noise to regression targets by explicitly passing noise_strength and noise_seed parameters in the command line. The noise seeding remains deterministic based on a stable dataset ordering for reproducible results.

Yes, the batch inference workflow supports GPU selection using cpu, cuda, or cuda:N formats. If you request CUDA but no GPU is available, it raises an error directly instead of silently downgrading to CPU.

The best way to verify the pipeline is using the smoke command. The Skill writes a lightweight smoke command and a full command in RUN.md, allowing you to validate CSV logging and pipeline functionality before executing the complete dataset iteration.

CSV outputs require a strict column order with the expression column fixed at the end. If the expression column is missing, your existing single-dataset script may not be correctly extracting and passing the required expression data to the CSV writer.