MIND-small · seeds 42, 43, 44 · RTX 4080 SUPER

Keep the notepad, 3 seeds

MIND
3 seeds
validation
RTX 4080 SUPER
keep the pad

Single-seed evidence is development. Three seeds make it confirmed. We re-ran the keep-the-pad versus wipe-and-rebuild protocol at two new training seeds (43 and 44) on the same MIND-small split as seed 42. Dataset split, snapshot, and notepad config were unchanged. Both new seeds kept the notepad ahead of wiping it. The top-1 ranking after keeping the pad matched a fresh rebuild on roughly 90% of scored pairs.

Takeaway

Keep beats wipe on all three seeds. Best NDCG@10: 0.3951, 0.3962, 0.3939 vs wipe 0.3885, 0.3894, 0.3897.

Best keep vs wipe · NDCG@10

Wipe, seed 42

0.3885

Keep, seed 42

0.3951

Wipe, seed 43

0.3894

Keep, seed 43

0.3962

Wipe, seed 44

0.3897

Keep, seed 44

0.3939

Summary · all seeds

Seed	Wipe NDCG@10	Keep NDCG@10	Best epoch	Top-1 agree
42	0.3885	0.3951	4	90.5%
43	0.3894	0.3962	6	89.5%
44	0.3897	0.3939	3	90.5%

Seed 43 · keep NDCG@10 by epoch

Epoch	Keep the pad	Same weights, rebuild eval
1	0.3842	0.3817
2	0.3921	0.3887
3	0.3889	0.3862
4	0.3920	0.3886
5	0.3956	0.3913
6	0.3962	0.3930
7	0.3937	0.3902
8	0.3947	0.3922
9	0.3924	0.3902

Seed 44 · keep NDCG@10 by epoch

Epoch	Keep the pad	Same weights, rebuild eval
1	0.3915	0.3876
2	0.3903	0.3866
3	0.3939	0.3892
4	0.3927	0.3890
5	0.3934	0.3891
6	0.3936	0.3898

Agreement · keep vs fresh rebuild

Seed	Pairs	Top-1 match	Largest score gap
42	200	90.5%	2.55
43	200	89.5%	2.73
44	200	90.5%	2.19

Isolated trainer copy; seed-42 files untouched. Dataset seed 42 for all three runs (same split). Only training seed changes. RTX 4080 SUPER, torch 2.13, CUDA 13. Seeds 43 and 44 ran in parallel on the same GPU. Wipe ran 10 epochs for all seeds. Keep early-stopped with patience 3. Validation only. Test split sealed.