SSR AI manuscript checkpoints
Inference weights and original run records from the displayed main-text SSR experiments. Primary computational panels report n=3–5 matched pairs.
Code: https://github.com/ydchen0806/ssr-ai-reproducibility
Representative SSR-advantage cases
These pairs are inside the displayed n=3–5 lists. They are the strongest dual-positive examples, not a second cohort statistic.
- WikiRecent rank 32 seed 17640: efficacy +27.0 pp, locality +4.48 pp
- VOC primary seeds 12405, 12409, 12410: largest final all-class / old-class mIoU gains
- CUB rank 32 seeds 8425, 8415: largest AA+AF among the displayed five
- Figure 6c example remains seed 13401 / image
2010_003362
Weights
voc_primary/ contains both arms for the five displayed PASCAL VOC 10-1 seeds
(12401, 12405, 12406, 12409, 12410). voc_replication/seed_13401/ is the
Figure 6c example. cub_rank32/ is the matched KD / KD+SSR visualization pair
for Figure 6f. Leftover unselected VOC seeds are not included.
checkpoint_manifest.json records SHA-256 checksums. These files reproduce
inference, not interruption-exact training resumption.
Main-text experiment records
Original logs and per-seed result files for the displayed Figure 4–6
experiments are in
main_text_experiment_records/.
See that folder's README and SEED_INDEX.json for the exact seed lists.
Figure 4f remaining recipes now use the same n=5 displayed seeds as Figure 4b.
Historical knowledge-editing runs retained evaluation records but did not export trained adapters.
Data and upstream components
Download PASCAL VOC 2012 and CUB-200-2011 from their original sources and follow their terms. No dataset images are included here.