Finetune on Spring dataset questions

Hello, thank you for releasing the training code. :)

I have a few questions. I want to finetune the model on an existing dataset that you noted in your supplementary materials, on the spring dataset. So I did a comparison with the stock version of the model and I obtained some metrics.
Note: The "test" spring dataset was obtained from the hugging face preprocessed datasets you mentioned in the eval section.

After that I took the entire train Spring dataset, and crafted the structure required to finetune the model on the Spring dataset, I followed your advice on the learning rates being 1e-5 for the decoder and 1e-6 for the encoder, and finetuned the model for an exact 1000 iterations ( too low?, it takes a while for a bigger number of iterations )
After doing so, I again reevaluated the new model with the new weights and observed my results are now slightly deviated from the original weights evaluation, being a little worse.

Questions:
Regarding the dataset crafting, in the meta.json structure, I also rescalled the intrinsics from:

{
"intrinsics": [
[
2181.8181818181815,
0,
960.0
],
[
0,
2181.8181818181815,
540.0
],
[
0,
0,
1
]
]
}

To:
{
"intrinsics": [
[
1.1363636363636362,
0,
0.5
],
[
0,
2.02020202020202,
0.5
],
[
0,
0,
1
]
]
}
, having cx and cy to 0.5

Is this normalized version of the intrinsics right?
Are there any tips in finetuning such a large foundation model?
Increasing the number of iterations will also play a huge part here?
Also the weight here, if we have only a dataset plays a part?
"name": "dataset_spring",
"path": "......",
"label_type": "synthetic",
"index": ".index.txt",
"depth": "depth.png",
"weight": 10, <- what role plays this when we have only one dataset? I think it doesn't matter?
"center_augmentation": 0.25,
"fov_range_absolute": [30, 150],
"fov_range_relative": [0.5, 1.0],
"image_augmentation": ["jittering", "jpeg_loss", "blurring", "shot_noise"]

Start with the supplementary materials and eval section, then inspect the meta.json structure and the Hugging Face preprocessed Spring test dataset used for comparison. Reproduce the stock-versus-finetuned evaluation using the stated 1,000 iterations and recorded intrinsics, and document answers about intrinsic normalization, iteration count, and the single-dataset weight.