MIND keep the notepad
Step one rebuilt a blank notepad from the last twenty clicks at every score. Here we ask whether training and scoring should instead keep that notepad, the way a live recommender would. Both rows use the same snapshot plus notepad on MIND-small, seed 42, with write-time features. One row wipes and rebuilds. The other keeps the pad, updates only new events, and does not reset memory. Keep-the-pad peaked at epoch four and then early-stopped. It beat wipe-and-rebuild on validation NDCG. A later agreement check crashed, so we did not write the persistent checkpoint. Test split stayed sealed.
Keeping the notepad beat wiping it. Best NDCG@10 is 0.3950 versus 0.3883.
NDCG@10 · seed 42 validation
Selected scores
| Row | AUC | NDCG@10 |
|---|---|---|
| Wipe and rebuild | 0.6251 | 0.3883 |
| Keep the notepad · epoch 4 | 0.6302 | 0.3950 |
Keep-the-pad NDCG@10 by epoch
| Epoch | Keep the pad | Same weights, rebuild eval |
|---|---|---|
| 1 | 0.3895 | 0.3857 |
| 2 | 0.3910 | 0.3873 |
| 3 | 0.3934 | 0.3889 |
| 4 | 0.3950 | 0.3905 |
| 5 | 0.3945 | 0.3909 |
| 6 | 0.3932 | 0.3900 |
| 7 | 0.3929 | 0.3898 |
Isolated copy of the trainer; original algorithm files were not edited. One seed, validation only, Modal L40S. Wipe-and-rebuild finished all ten epochs. Keep-the-pad early-stopped after epoch 7 (patience 3). An agreement check then crashed on a missing batch field, so the persistent checkpoint was not written. Ranking numbers are from the epoch logs. Test split sealed. Layout of snapshot, notepad, and write network is on the method page, not this run.