Output format problem about RoboBrian2.0-32B

[lizj18]

Hi, first thanks for the open-source of such great work.

I met some format problem when using your original format rule to extract the "think" and "answer" from the output of 32B version model when enable "thinking" mode. Here are the case:
Case 1:

##### RAW_OUTPUT #####
based on the consistent visual evidence of the window and fridge being part of the scene from this angle, the answer is:
<answer>B. Window and fridge</answer></think><answer><think></think><answer>B. Window and fridge</answer></answer>
##### ANSWER #####
<think>
#################

Case 2：

##### RAW_OUTPUT #####
Considering the spatial layout and the task goal, the most logical next step is to initiate a left turn to reposition myself appropriately for exiting the bedroom and proceeding towards the hallway.</think><answer><think></think><answer>left 90</answer></answer>
##### ANSWER #####
<think>
#################

Case 3:

##### RAW_OUTPUT #####
Yes</think>
##### ANSWER #####
(empty)
#################

As you can see the raw outputs of 3 cases seems not in desired format which lead to the wrong extracted answers(I leverage the official method from inference.py to extract "answer" and "thinking"). And I didn't find such problem in the 7B version.

I wanna ask if there are any solution for such problem or did I do some think wrong?

[cmyopi]

Thank you very much for your question. Could you please provide the steps to reproduce your issue, including your question and the images input into the model?