nemo_gym.utils.path_utils
nemo_gym.utils.path_utils
Module Contents
Functions
Data
API
results/rollouts.jsonl -> results/rollouts_aggregate_metrics.json.
Mirrors how rollout collection and reverification name the file they write, so consumers
(e.g. gym eval compare) derive the same path the writers produced.
results/rollouts.jsonl -> results/rollouts_sources.json: the shards gym eval aggregate merged into it.
Where gym eval resubmit --recover --extend-attempts N records the attempts it added to every rollout’s cap.
results/math.jsonl -> results/math_batch_status.json.
results/rollouts.jsonl, mlflow -> results/rollouts_mlflow_export_done: set once the run was exported.
results/rollouts.jsonl, mlflow -> results/rollouts_mlflow_run_id: the run a live exporter logs to.
results/rollouts.jsonl -> results/rollouts_materialized_inputs.jsonl.
Where gym eval resubmit --recover --rerun-masked-without-artifact journals the operator’s approval to run again
masked results that kept a saved answer but have neither a gradeable artifact nor a grade: beside the rollouts file
and named after it, like the attempt extension, so two rollouts files in one directory never share approvals.
results/rollouts.jsonl -> results/rollouts_rerun_masked_journal.jsonl.
results/rollouts.jsonl -> results/rollouts_resolved_config.yaml.
Where gym eval resubmit records the failures sidecar’s line count; failures before it no longer count.
results/rollouts.jsonl -> results/rollouts_run_summary.json.
The per-run summary a Harbor run writes beside its rollouts (nemo_gym.data.tasks.harbor.run_summary),
named like the other sidecars so a renamed --output keeps them together.