NVIDIA-NeMo / NVIDIA-NeMo/RL

llm_grpo_llama3_2_1b_instruct_1n8g_megatron_generation nightly test reward not stable

Open
#3,257 0 comments 0 reactions 1 assignee Assigned to @terrykong View on GitHub
accuracy bug community-request waiting-on-maintainers
Dominant language
Python
Stars
2k
Forks
561
Avg merge
4d 5h
Merged PRs (30d)
145

Description

**Describe the bug**

Final step train reward for this test recipe is not always stable

**Steps/Code to reproduce bug**

Tests off the main branch.

Historical details
Image

Another example was observed durign the v0.7.0 release testing:

```
Expected data["train/reward"]["500"] > 0.1
Value: 0.05
```

Please help make the nightly test stable.

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.