[arendu]

What does this PR do?

Adds a raw_rewards_and_lengths field to each per-task (group-level) row in the
reward profiling output produced by RewardProfiler.profile_from_data.

The field contains one entry per rollout, sorted by rollout index, with
_ng_rollout_index, reward, and input_tokens/output_tokens/total_tokens
(token keys included only when present in the rollout's usage). It is a cheap
projection of the existing rollout_infos, so it flows through unchanged to both
the *_reward_profiling.jsonl artifact and the group-level metrics returned by
compute_aggregate_metrics.

Motivation: aggregate stats like pass rate discard the per-rollout evidence
behind them. Preserving each rollout's reward and length in the profiled output
enables downstream length-aware filtering, checking whether passing rollouts are
systematically shorter/longer than failing ones, and re-deriving pass rates
(passed = count(reward > 0), total = len(list)) without re-running rollouts.

Changes (+25 lines, no deletions):

• nemo_gym/reward_profile.py: build raw_rewards_and_lengths from the sorted
  rollout_infos in profile_from_data (+8)
• tests/unit_tests/test_reward_profile.py: assert the new field's exact
  per-rollout contents in test_rollout_infos_are_sorted_and_pass_rate_is_recoverable,
  and add the key to the _clean_metrics ignore set (+17)

Checklist

• I have read the contributing guidelines.
• The change is focused; unrelated "drive-by" edits are tracked as separate issues/PRs.
• Tests added or updated and pass locally, or N/A for docs-only / non-code changes (so CI unit/server checks pass when applicable).
• Pre-commit checks pass locally (pre-commit run --all-files) (so CI lint/format/copyright pass).
• All commits have DCO sign-off (git commit -s) (so the DCO check passes).

[Copilot]

Pull request overview

This PR enhances NeMo Gym’s reward profiling outputs by preserving per-rollout evidence (reward and token lengths) alongside aggregate metrics, enabling downstream analyses (e.g., length-aware filtering and pass-rate re-derivation) without re-running rollouts.

Changes:

• Add a raw_rewards_and_lengths field to each per-task (group-level) metrics row produced by RewardProfiler.profile_from_data, derived from the already-sorted rollout_infos.
• Update unit tests to (1) assert the exact contents of the new field in the “sorted rollout infos” test and (2) ignore the new key in _clean_metrics where appropriate.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
nemo_gym/reward_profile.py	Adds raw_rewards_and_lengths as a lightweight projection of sorted rollout_infos into group-level metrics rows.
tests/unit_tests/test_reward_profile.py	Updates metrics-cleaning to ignore the new key and adds a precise assertion for raw_rewards_and_lengths contents/order.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[cmunley1]

Doesnt the output already contain the fields in raw_rewards_and_lengths, for example from https://github.com/NVIDIA-NeMo/Gym/blob/main/resources_servers/longmt_eval/data/example_rollouts_reward_profiling.jsonl#L1C1-L2C1

    {
      "rollout_id": "28:0",
      "_ng_task_index": 28,
      "_ng_rollout_index": 0,
      "reward": 0.7699517591132058,
      "input_tokens": 3690,
      "output_tokens": 4182,
      "total_tokens": 7872,
      "comet_qe": 0.7699517591132058,
      "lang_fidelity": 1.0,
      "total_seg": 108,
      "misaligned_seg": 1
    }