A hopelessly try to replicate the YOLOv3 kernel

[AlexeyDate]

I have already created a section in the issue dedicated to my problem. At the moment I want to appeal to everyone interested in yolov3.
I will be glad and very grateful for any help!

My question will not be directly related to your repository. I'm sorry that I have to come here.
Throughout the summer, I tried to implement yolov3 based on box loss. A long time later, I got minimal results, which are only a small part of my success. I have researched your repository in detail and highlighted the key ideas for my implementation. Unfortunately, now my ideas in finding the error are completely exhausted and I hope for minimal hints on the way to a solution.

I conduct all experiments based on the following dataset: https://www.kaggle.com/datasets/biancaferreira/african-wildlife
In my attempt to implement yolov2, the mAP results were about 0.7, but now they barely reach 0.3. The classification loss is very different between the training and validation sample - I tried to make a visual screenshot of my training. I studied all hyperparameters and detailed points for a long time and did not find any improvement in this.
Could someone please take a look at the key points in the lost file, data upload error and others?

I have lost hope in solving this problem, so I am writing to community.

My repo: https://github.com/AlexeyDate/YOLOv3-test-
Full dataset in the desired: https://drive.google.com/drive/folders/1uSaz_23A5fWbFByHcYwYh6wHzlArBqtv?usp=sharing
Pretrained weights: https://pjreddie.com/media/files/darknet53_448.weights

Example 1

Example 2

With many thanks to your efforts and the repository!