MIND keep the notepad, confirmation
This is a second isolated copy of the keep-the-pad trainer, run on a university RTX 4080 SUPER instead of Modal. Same MIND-small snapshot plus notepad, seed 42, write-time features. One row wipes the pad and rebuilds from the last twenty clicks. The other keeps the pad across visits. Keep-the-pad peaked at epoch four, then early-stopped after epoch seven. It beat wipe-and-rebuild on validation NDCG. This time we wrote the persistent checkpoint, and a two-hundred-pair score check agreed on the top item for most impressions. Test split stayed sealed.
Keeping the notepad beat wiping it, and the run finished. Best NDCG@10 is 0.3951 versus 0.3885.
NDCG@10 · seed 42 validation
Selected scores
| Row | AUC | NDCG@10 |
|---|---|---|
| Wipe and rebuild · epoch 8 | 0.6252 | 0.3885 |
| Keep the notepad · epoch 4 | 0.6303 | 0.3951 |
Keep-the-pad NDCG@10 by epoch
| Epoch | Keep the pad | Same weights, rebuild eval |
|---|---|---|
| 1 | 0.3893 | 0.3856 |
| 2 | 0.3912 | 0.3873 |
| 3 | 0.3934 | 0.3888 |
| 4 | 0.3951 | 0.3907 |
| 5 | 0.3947 | 0.3910 |
| 6 | 0.3933 | 0.3902 |
| 7 | 0.3932 | 0.3900 |
Keep versus wipe scores
| Check | Value |
|---|---|
| Pairs compared | 200 |
| Same top item | 0.905 |
| Largest absolute score gap | 2.55 |
Isolated copy of the trainer; original algorithm files and the Modal keep-the-pad folder were not edited. One seed, validation only, RTX 4080 SUPER, torch 2.13 CUDA 13. Wipe-and-rebuild finished all ten epochs (best at epoch 8). Keep-the-pad early-stopped after epoch 7 (patience 3). Persistent checkpoint was written. Embeddings were recomputed on this box with the same Qwen script and revision as the Modal volume. Test split sealed. Layout of snapshot, notepad, and write network is on the method page.