nemo-gym-reward-profiling

What problem does it solve?

Reward profiling in Nemo Gym can be hard to run correctly and interpret when repeats, materialized inputs, and per-rollout results must be joined by task and rollout identity.

Core Features & Use Cases

• Orchestrates the standard profiling flow: use ng_run to start servers, ng_collect_rollouts to produce rollouts.jsonl and materialized inputs, and ng_reward_profile to generate reward profiling outputs.
• Explains how identities and artifacts connect: clarifies how _ng_task_index and _ng_rollout_index link *_materialized_inputs.jsonl, rollouts.jsonl, and *_reward_profiling.jsonl.
• Supports practical profiling scenarios: repeated rollouts for per-task averages and variance, strict vs partial profiling (including allow_partial_rollouts), and inspecting rollout_infos for reward and token metrics.

Quick Start

Run ng_run to start the Nemo Gym services, then ng_collect_rollouts to write rollouts.jsonl and *_materialized_inputs.jsonl, and finally ng_reward_profile to produce *_reward_profiling.jsonl from the materialized inputs and rollout artifacts.