Pretrained seq2seq_model_32bits.pth seems corrupted in both folders

[FHCzzs]

Hi! I was trying to reproduce DynamicRTL and ran into an issue with the pretrained model files in the repo.

There are two copies of seq2seq_model_32bits.pth:

Path	Size (bytes)	torch.load
src/gdrtl_arch/seq2seq_model_32bits.pth	69092401	failed
src/test_gdrtl_arch/seq2seq_model_32bits.pth	69092364	failed

Both files are almost the same size, and both fail to load with:

RuntimeError: PytorchStreamReader failed reading zip archive: invalid header or archive is corrupted

These are the sizes from a fresh clone; we didn’t overwrite the files.

The training code actually loads the one under gdrtl_arch (same directory as encoder_decoder.py). The copy under test_gdrtl_arch looks more like a backup or test version.

I couldn’t find any other pretrained checkpoints in the repo, so I wanted to check:

• Were these .pth files incomplete when uploaded?
• Is there an original, non-corrupted checkpoint available (e.g. Google Drive / release asset)?
• If the original file isn’t available, is there a recommended way to retrain the Seq2Seq encoder?

A working checkpoint would really help with reproducing the results.
Thanks!