The difference ranking result in the python and libcatboost by max_relevance times

[akalend]

Problem:
The difference result of the rank model in the python and libcatboost
Разные результаты при работе модели в python и си кода с использованием libcatboost. Отличаются на величину max_relevance

Steps of problem:

1. I use the training model from tutorial https://github.com/catboost/tutorials/tree/master/ranking

2. The url of my notebook of the model is https://colab.research.google.com/drive/1IXzicNub6CtzPRW3ZUfFkud48EDgaLhC?usp=sharing

3. The RMSE criterion was used.

4. The model saved as cbm format and use in the C-code. The example of code : https://github.com/cyrusmsk/tutorials/tree/clang_dlang_applier_tutorials/apply_model/clang

5. compare results from Python and C-code (libcatboost) : the results differ by max_relevance times (4 times) . See cell number 29 in the notebook (Itemt 2).

The original C-code https://github.com/akalend/pg_ml/blob/main/ml.c

catboost version: catboost-1.2.3-cp310-cp310-manylinux2014_x86_64.whl ( Google collab)
Operating System: Ubuntu 20.0
CPU: iCore i5 local machine

[andrey-khropov]

I cannot reproduce this.

See the training code in python and inference code in C++. In my example they produce the same predictions.
I tried both single-object and multi-objects variants.