Fix RLVR integ test

[zhaoqizqwang]

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Claude finished @zhaoqizqwang's task in 43s —— View job

---

PR Review: Fix RLVR integ test

• Read the PR diff (/tmp/pr.diff)
• Review changed test file and surrounding context
• Verify against the code being tested

Review

This is a correct, minimal fix. ✅

The two RLVR dry-run integ tests (test_rlvr_dry_run_returns_none and test_rlvr_serverful_dry_run_returns_none) construct an RLVRTrainer without any reward signal. Since RLVRTrainer.train() now validates that a reward signal is present (rlvr_trainer.py:417-424), calling train(dry_run=True) raises ValueError before reaching the return None path, so both tests would fail.

Adding trainer.hyperparameters.preset_reward_function = "prime_code" satisfies that validation and lets the dry run complete as expected.

Consistency check: the change matches the exact pattern already used in test_rlvr_trainer_integration.py:96 and the value/attribute suggested by the trainer's own error message (rlvr_trainer.py:423). "prime_code" is a documented preset value, and the attribute name is correct.

Minor observations (non-blocking):

• The reward-signal requirement is intrinsic to RLVRTrainer, so these tests would be more robust setting preset_reward_function at (or immediately after) construction is fine as-is — mutating trainer.hyperparameters post-init is already the established convention in this suite, so no change needed.

No correctness, security, or performance concerns. LGTM.

[mujtaba1747]

nice