Failed to parse the scorer response

[zqxie77]

Description

Hi, I am trying to reproduce the CoMAS training process using Qwen2.5-3B-Instruct as the base model. The environment setup (Ray, VLLM, DeepSpeed) is successful, and the model weights are loaded correctly.

However, during the Simulating phase (rollout), I am encountering frequent parsing errors related to the scorer's response. It seems the scorer agent is either returning None or generating text that cannot be parsed as an integer.

Error Logs

Here are the specific error logs appearing in the console during Simulating round:

Failed to parse the scorer response: 'NoneType' object has no attribute 'text'
Failed to parse the scorer response: 'NoneType' object has no attribute 'text'
Failed to parse the scorer response: invalid literal for int() with base 10: '" and "'
Failed to parse the scorer response: 'NoneType' object has no attribute 'text'
Simulating round 2 out of 2...

[xxyQwQ]

Hi, thank you for your interests in our work.

Generating contents with incorrect formats is a normal case since Qwen2.5-3B-Instruct is not quite a strong model. We have applied format rewards to solve this problem. You will see fewer parser failures with the training process going on. These error logs will not affect the overall training.