Can you please share more metrics for the Qwen3 30B A3B training?

[Shangda-Yang]

Feature proposal

I am trying to run the Qwen3 30B A3B example provided and would like to know more about the trends of entropy, loss, response length etc.. Can you please share more detailed metrics for this example? Thank you.

Motivation and use case

Run the Qwen3 30B A3B example

Related area

docs / examples

Proposed design or API

No response

Alternatives considered

No response

Additional context

No response

Contribution

• I am willing to help implement this feature.
• I am willing to help test or validate this feature.

[yyDing1]

Thanks for your attention. Below is the training curve of an old version. (Qwen3-Coder-30B-A3B-Ins)

We have refactored the framework, and new results will be presented in the docs soon. Benefit from optimized harness engineering, the train-time reward could be higher than the old one.

[yyDing1]

Hi, results and relevant details has been updated in the document: https://uni-agent.readthedocs.io/en/latest/quickstart/rl-training.html#results

[Shangda-Yang]

Thanks for sharing more details. That's super helpful.