Running metrics offline on a folder of structures

You do not have to re-run generation to get metrics. evaluate.py scores a folder of structures you already have, on the same Hydra config, so metrics_tags and any custom metric config work exactly as above.

rfd4-evaluate \
    inputs=/path/to/structures \
    metrics_tags='[unconditional_refolding]' \
    out_dir=/path/to/eval_output

evaluate.py scans inputs recursively for CIF-like files. out_dir is required, because evaluate.py has no run directory to derive one from, and it gets the usual evaluation_metrics/ subdirectory.

Pass task_csv=<csv with designed_structure_path and task columns> when different structures need different tags or pass criteria. Without it, every structure is scored the same way. The Developer Guide points at this same script for custom metric work, so you write a metric once and use it both inline during generation and offline over old outputs.

For large batches split across SLURM array jobs (launch_generate.sh --njobs N), each job writes its own per_example_metrics_job{N}.csv. Concatenate them afterwards.

python -m rfproteina.inference_utils.aggregate_metrics \
    ./runs/my_run/inference_metrics/

launch_generate.sh also runs this for you as a dependent job. See Evaluating results for FoldSeek diversity clustering and the other aggregate_metrics options, and the output format reference for the exact file layout these commands produce.