Machine Translation on Flores-101 (val test)
46.8CHRFLSL-NAS
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LSL-NASdense pre-training=true, Total Parameters (|θ|)=441M, Effective Parameters (|θeff|)=186M2023.05 | 46.8 | 27.5 | |
| Separate Decoder Baselinehidden dim=768, Total Parameters (|θ|)=509M, Effective Parameters (|θeff|)=296M2023.05 | 46.5 | 27.3 | |
| LSL-NASTotal Parameters (|θ|)=441M, Effective Parameters (|θeff|)=186M2023.05 | 46.4 | 27.2 | |
| LSLSRC layers={1, 2}, TGT layers={15, 16}, Total Parameters (|θ|)=413M, Effective Parameters (|θeff|)=186M2023.05 | 46.3 | 27 | |
| Separate Decoder Baselinehidden dim=704, Total Parameters (|θ|)=453M, Effective Parameters (|θeff|)=268M2023.05 | 46.2 | 26.9 | |
| LSLSRC layers={1, 2, 3}, TGT layers={14, 15, 16}, Total Parameters (|θ|)=470M, Effective Parameters (|θeff|)=186M2023.05 | 46.2 | 27 | |
| Language Adapters ENC (SRC+TGT)bottleneck dimension=256, Total Parameters (|θ|)=342M, Effective Parameters (|θeff|)=228M2023.05 | 46.1 | 26.7 | |
| LSLSRC layers={3, 4}, TGT layers=0, Total Parameters (|θ|)=356M, Effective Parameters (|θeff|)=186M2023.05 | 46.1 | 26.7 | |
| Language Adapters ENC (SRC+TGT)bottleneck dimension=512, Total Parameters (|θ|)=384M, Effective Parameters (|θeff|)=270M2023.05 | 46 | 26.7 | |
| Separate Decoder Baselinehidden dim=640, Total Parameters (|θ|)=399M, Effective Parameters (|θeff|)=240M2023.05 | 45.9 | 26.6 | |
| Language Adapters ENCbottleneck dimension=128, Total Parameters (|θ|)=321M, Effective Parameters (|θeff|)=207M2023.05 | 45.8 | 26.4 | |
| LSLSRC layers=Ø, TGT layers={13, 14, 15}, Total Parameters (|θ|)=385M, Effective Parameters (|θeff|)=186M2023.05 | 45.8 | 26.5 | |
| Language Adapters ENCbottleneck dimension=256, Total Parameters (|θ|)=342M, Effective Parameters (|θeff|)=228M2023.05 | 45.7 | 26.3 | |
| Language Adapters ENC (SRC+TGT)bottleneck dimension=128, Total Parameters (|θ|)=321M, Effective Parameters (|θeff|)=207M2023.05 | 45.7 | 26.3 | |
| Language Adapters ENCbottleneck dimension=512, Total Parameters (|θ|)=384M, Effective Parameters (|θeff|)=270M2023.05 | 45.6 | 26.3 | |
| Separate Decoder BaselineTotal Parameters (|θ|)=299M, Effective Parameters (|θeff|)=186M2023.05 | 45.5 | 26 | |
| LSLSRC layers={13, 14, 15}, TGT layers={3, 4}, Total Parameters (|θ|)=441M, Effective Parameters (|θeff|)=186M2023.05 | 45.2 | 25.7 |