I could not really find a satisfying answer here yet, this question touches it but kind of reversed.
For example the training framework for Yolo v7 is used to train a neural network using the config (cfg) file from the repo which is licensed under GPL-3 using custom data.
Now the model gets exported to onnx (or tflite or any other major inference runtime) and used with inference scripts that are completely self-written (inference done by onnxruntime, especially post-processing like decoding and nms self-written).
Is (2) now derivative work and must be distributed under GPL-3 too by using the neural network weights?