Using Poisson objective and a baseline gives negative predictions

Problem: Using Poisson objective and a baseline gives negative predictions
catboost version: 0.23.2
Operating System: Linux, redhat enterprise 7
CPU: 40 * Intel Corporation Xeon E7 v3/Xeon E5 v3/Core i7
GPU: No

Hello,
I am trying to predict an integer target using Poisson objective. I need an offset so I'm using the set_baseline() function on the Pools. However, the prediction given by Catboost is very bad and when setting the baseline with a log gives me negative predictions.
Am I missing something ? Or Could it be a problem similar to the issue #1184 but for the baseline ?

Thank you for your help !