Why output safety assessment before categories？ #1
Description
Hello, I noticed that the output format of your model is a bit unusual.
You have the model output the safety assessment result first, followed by the category and the explanation.
This means that when the model makes its judgment, it doesn’t actually see the "Explain" content (which can also be understood as "think") beforehand. As a result, the model is likely making a direct classification with only a short reasoning process, and then generating the explanation based on the classification result.
This approach might not fully leverage the "think" part.
I also looked at a similar work—GuardReasoner—which emphasizes the importance of thinking first.
May I ask why you chose to classify first and then explain? Or have you conducted any ablation studies where the model thinks first and then classifies?