The code file for evaluating math and gsm8k is missing

[zaiquanyang]

Can you upload these files ?

[JingyangOu]

Thanks for the issue!

For GSM8K and Math evaluations, we did not include separate evaluation code in this repo. Instead, we directly use the official evaluation scripts provided by the respective model codebases, as described in the README.

[zaiquanyang]

thanks for your response~
For GSM8K and Math evaluations, I also notice that the works like SPG and d1 adopt differerent evaluation instead of using the opencompass in LLaDA. Could you briefly describe the differences between the two?

[JingyangOu]

Thanks for the clarification!

d1 is an early work on dLLMs with RL, and at that time the official LLaDA-Instruct evaluation code was not yet available. As a result, d1 implemented its own evaluation framework with a custom system prompt. SPG, although a concurrent work with ours, chose to directly follow the evaluation setup introduced in D1.

In contrast, we adopted the official LLaDA evaluation code once it became available, as it allows us align the reported LLaDA and LLaDA-1.5 results. The datasets and metrics remain the same; the difference mainly comes from the prompt.