These are the exact lists every number in the paper was produced on. Only the EO side is stored; the SAR path is derived from it, so the two can never drift apart:
| Dataset | Rule |
|---|---|
| QXS-SAROPT | opt_256_oc_0.2/ → sar_256_oc_0.2/ |
| SAR2Opt | the EO directory's last character is replaced by A (testB → testA) |
| List | Items |
|---|---|
qxs-saropt/train_eo_list.txt |
16,001 |
qxs-saropt/test_eo_list.txt |
3,999 |
sar2opt/train_eo_list.txt |
1,450 |
sar2opt/test_eo_list.txt |
627 |
Paths are relative to the dataset root. No imagery is included.
There is no held-out validation split. Training monitors a sample drawn from the training data; it is used to watch the loss, never for model selection. The released checkpoints are the last step of a fixed-length schedule, not a best-of.