Questions about implementation details for reproduction (feature extractor, dims, seeds, hyperparameters)

[HLL-hll666]

Hi, thank you for the great work and for open-sourcing the code.

I'm trying to reproduce the results in the paper and I have a few questions about implementation details that are not fully specified in the paper or code. I would really appreciate your clarification.

1. Feature extractor
   Which exact model was used to extract patch features? Is it UNI (ViT-L/16)? Or another pathology foundation model (e.g., CONCH, CTransPath)? Which weights/version? Was the extractor frozen during training?

2. Feature dimension and projection
   If UNI is used, its output is 1024-dim. Table 3 says the MIL input is 128-dim. Is 128 the raw feature dim or after projection? If projection, is it a linear layer or MLP?

3. Random seeds
   You mentioned each experiment was run 5 times with 5 different seeds. Could you specify which 5 seeds?

4. Training hyperparameters
   Could you share: number of epochs, learning rate, optimizer, weight decay, dropout, batch size (for feature extraction and MIL training separately)?

5. Data splits
   For WSI classification, are train/val/test splits randomly generated per seed or fixed? For survival analysis, could you share the 5-fold split files?

6. Model configuration
   Number of 2DMamba layers U, Mamba inner dimension, and whether mamba_2d_pad_token is set to "trainable" (as in Table 6)?

7. Patch preprocessing
   I saw in issue patch size for extract features using CLAM #13 that patches are 512x512 at 20x (level=1) using CLAM. Is there any additional downsampling or resizing before feeding patches into the feature extractor? Since UNI takes 224x224 input, how is the 512x512 patch handled?

Thank you very much for your time.