nemo-rl-auto-research

Automate NeMo-RL experiments by orchestrating baseline setup, branch-based hypothesis trials, and per-experiment TSV logging. It reads target recipes, enforces branching discipline, and maintains a complete audit trail of campaign decisions for full reproducibility.

Manage iterative reinforcement learning trials by creating a branch per hypothesis under a shared prefix and enforcing stop rules. Results are written to a durable TSV ledger recording metadata such as branch, commit, metric, memory, and log paths.

Long-running experiment campaigns are supported through a session memory that preserves state. This enables resume and handoff capabilities, while safety checks and per-experiment budgets ensure controlled execution across extended research runs.

Enforce git workflow discipline by creating a dedicated branch per hypothesis under a shared prefix for each experiment. The system checks prerequisites and records commit metadata, ensuring every iterative trial remains isolated and reproducible.

The experiment ledger uses TSV format to track NeMo-RL metrics and metadata. It records per-experiment results including branch names, commit hashes, metrics, memory state, and log paths to maintain a durable audit trail across campaigns.