Scaling Law Extrapolation Accuracy on Japanese-Indonesian-Swahili 1-stage (test)
0.95R² (C)He + Dual PL
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| He + Dual PL#P=7, training_regime=Multi-lingual only, epochs=single-epoch, test_set=1-stage test2024.10 | 0.95 | 0.72 | 0.89 | 0.87 | 0.82 | 0.85 | |
| M3 Scaling Law#P=8, training_regime=Multi-lingual only, epochs=single-epoch, test_set=1-stage test2024.10 | 0.95 | 0.85 | 0.85 | 0.76 | 0.82 | 0.85 | |
| M3 Scaling Law - Dual PL#P=7, training_regime=Multi-lingual only, epochs=single-epoch, test_set=1-stage test, dual_pl=false2024.10 | 0.63 | 0.42 | 0.63 | 0.12 | -0.16 | 0.33 | |
| Sedova#P=9, training_regime=Multi-lingual only, epochs=single-epoch, test_set=1-stage test2024.10 | 0.62 | 0.47 | 0.79 | 0.51 | -0.19 | 0.44 | |
| He#P=6, training_regime=Multi-lingual only, epochs=single-epoch, test_set=1-stage test2024.10 | 0.51 | 0.06 | 0.7 | 0.28 | -0.27 | 0.26 |