s3_e5
done · peak unseen_mismatch 71.3
α schedule
α 3.0 ×50 α 3.087 ×50 α 3.1739 ×50 α 3.2609 ×50 α 3.3478 ×50 α 3.4348 ×50 α 3.5217 ×50 α 3.6087 ×50 α 3.6957 ×50 α 3.7826 ×50 α 3.8696 ×50 α 3.9565 ×50 α 4.0435 ×50 α 4.1304 ×50 α 4.2174 ×50 α 4.3043 ×50 α 4.3913 ×50 α 4.4783 ×50 α 4.5652 ×50 α 4.6522 ×50 α 4.7391 ×50 α 4.8261 ×50 α 4.913 ×50 α 5.0 ×50
Figures
Results
accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 5.0
seen_* probed at reference α = 1.4 · unseen_* = held-out pairings, uniform (α-independent) · 1000 pairs per condition
| iteration | seen_match | seen_mismatch | unseen_match | unseen_mismatch |
|---|---|---|---|---|
| 50 | 51.3 ± 2.2 | 49.6 ± 0.8 | 50.0 ± 1.7 | 50.3 ± 1.6 |
| 100 | 56.6 ± 6.5 | 53.8 ± 4.1 | 53.8 ± 5.4 | 50.5 ± 2.0 |
| 150 | 76.4 ± 14.0 | 74.9 ± 12.2 | 67.9 ± 13.5 | 64.0 ± 11.6 |
| 200 | 87.7 ± 3.4 | 86.0 ± 3.8 | 75.3 ± 6.8 | 70.8 ± 6.4 |
| 250 | 88.4 ± 2.5 | 86.6 ± 1.9 | 73.7 ± 4.1 | 70.4 ± 3.7 |
| 300 | 89.9 ± 2.0 | 88.2 ± 1.2 | 74.4 ± 4.6 | 71.3 ± 4.5 |
| 350 | 89.8 ± 2.3 | 88.3 ± 2.2 | 73.5 ± 4.8 | 70.2 ± 3.3 |
| 400 | 89.4 ± 2.4 | 87.4 ± 2.0 | 72.1 ± 5.0 | 68.6 ± 3.6 |
| 450 | 90.0 ± 1.4 | 88.2 ± 1.6 | 73.6 ± 3.9 | 69.2 ± 2.7 |
| 500 | 90.1 ± 1.5 | 88.5 ± 1.6 | 72.8 ± 3.2 | 68.9 ± 2.3 |
| 550 | 90.9 ± 1.3 | 89.2 ± 1.7 | 75.1 ± 4.3 | 70.7 ± 4.4 |
| 600 | 90.3 ± 1.4 | 88.1 ± 2.0 | 73.9 ± 4.3 | 68.1 ± 4.3 |
| 650 | 90.8 ± 1.4 | 88.5 ± 1.4 | 73.8 ± 5.7 | 69.0 ± 4.8 |
| 700 | 90.1 ± 1.5 | 87.9 ± 2.2 | 73.1 ± 3.0 | 68.8 ± 4.2 |
| 750 | 90.1 ± 1.2 | 87.5 ± 1.5 | 72.8 ± 3.1 | 68.6 ± 3.3 |
| 800 | 89.7 ± 1.2 | 87.6 ± 1.5 | 73.1 ± 2.9 | 69.1 ± 3.1 |
| 850 | 90.1 ± 1.4 | 88.1 ± 1.7 | 72.8 ± 4.6 | 68.3 ± 3.7 |
| 900 | 89.5 ± 1.5 | 87.5 ± 1.8 | 72.9 ± 3.5 | 69.1 ± 3.4 |
| 950 | 89.8 ± 1.4 | 87.6 ± 1.2 | 73.2 ± 3.8 | 68.5 ± 3.8 |
| 1000 | 89.6 ± 1.8 | 87.5 ± 2.4 | 73.1 ± 3.8 | 68.2 ± 3.9 |
| 1050 | 89.3 ± 1.5 | 87.4 ± 2.0 | 72.2 ± 4.1 | 67.0 ± 4.3 |
| 1100 | 89.5 ± 1.0 | 88.0 ± 1.6 | 74.2 ± 3.1 | 69.2 ± 2.8 |
| 1150 | 89.7 ± 1.1 | 87.4 ± 1.9 | 74.0 ± 3.9 | 68.8 ± 4.0 |
| 1200 | 89.8 ± 1.2 | 87.8 ± 1.2 | 72.0 ± 3.4 | 68.4 ± 2.7 |