docs/reproducibility.md

Reproducibility Notes

Checkpoint Conversion

The released files were converted from training checkpoints with:

ckpt = torch.load("final.pt", map_location="cpu")
state = {k.removeprefix("module."): v.detach().cpu().contiguous() for k, v in ckpt["model"].items()}
safetensors.torch.save_file(state, "model.safetensors")

The original .pt files contain optimizer and training state. The released .safetensors files contain only the model state dict required for inference.

Verification

Each converted checkpoint was reloaded with safetensors.torch.load_file, and the tensor key set was checked against the source PyTorch checkpoint. A forward pass sanity check was also run using the TeX-1500 codebase and a synthetic [1, 64, 64, 64] HSI tensor.

Data Access

The TeX-1500 dataset is gated on Hugging Face. Users need to log in and accept the dataset access conditions before downloading samples for reproduction.