auto-research

To automate NeMo-RL experiments for directed hypothesis testing, this Skill iteratively plans, runs, and evaluates runs while applying objective metrics with explicit stop rules. It orchestrates the entire lifecycle from baseline creation to final evaluation.

Reproducibility in NeMo-gym recipe workflows is ensured by using deliberate git branching per hypothesis, committing minimal changes, and logging authoritative metrics and command provenance in a durable TSV ledger.

Metric tracking during autonomous reinforcement learning campaigns works by extracting authoritative metrics like validation accuracy or reward, evaluating them against monitored stop rules, and recording outcomes in a TSV schema.

Yes, you can use uv for reproducible command execution in NeMo-RL experiment automation. The Skill executes iterative runs via uv to maintain strict command provenance and reproducible environments.

Experiment automation supports distributed environments by selecting the right launcher for your infrastructure, including local, Slurm, or Kubernetes, while preserving existing user work and applying operational guardrails.

You maintain session continuity across long reinforcement learning research campaigns using checkpointed session-memory handoffs. This preserves state and context across iterative planning, running, and evaluating experiments.