nemo-gym-reward-profiling

Nemo Gym Reward Profiling

Invocation Check

Use this skill when the user wants to run, understand, or lightly modify Nemo Gym reward profiling. Keep the answer oriented around the normal workflow:

gym env start starts model/resources servers, gym eval run --no-serve writes rollout artifacts, and gym eval profile generates profiling output from those artifacts.

If the user is primarily debugging a failed job or stack trace, use the nemo-gym-debugging skill first.

Basic Workflow

1. Identify the environment config paths and input JSONL.
2. Start Gym servers with gym env start.
3. Collect rollouts with gym eval run --no-serve; this writes rollouts.jsonl and *_materialized_inputs.jsonl.
4. Run gym eval profile on the materialized inputs and rollout JSONL to generate *_reward_profiling.jsonl.
5. Inspect line counts and profile rows.

Repeated rollouts are the main profiling lever. num_repeats=1 is valid, but per-task averages and variance are only meaningful with multiple rollouts per task.

Core Concepts

• *_materialized_inputs.jsonl: expanded collection inputs after repeat expansion, agent defaults, and task/rollout id assignment.
• rollouts.jsonl: one completed rollout/result per materialized input row.
• *_reward_profiling.jsonl: one summarized profile row per original task with at least one completed rollout.
• _ng_task_index: original task/sample id.
• _ng_rollout_index: repeated rollout id for that task.
• rollout_infos: compact per-rollout info inside each task profile row, including reward, token usage, and numeric rollout metrics when available.

Keep reward-to-length or reward-to-token analysis keyed by both _ng_task_index and _ng_rollout_index.

Reference Loading

Load references only when the user needs that detail:

• Read references/quick-start.md for a generic command template and the minimal run sequence.
• Read references/output-format.md to explain materialized inputs, rollout JSONL, reward profile rows, rollout_infos, and partial profiling.

Practical Defaults

• Treat gym eval profile as the reward profiling step; rollout collection does not write reward profile files.
• Run strict profiling by default. If rollout collection stopped early, use ++allow_partial_rollouts=True to profile completed rollouts and drop original input rows with no completed rollout.
• Trust the target checkout's CLI help and nemo_gym/reward_profile.py over memory if flags differ.