List your company

Machine Translation on WMT en-de

30.17BLEUdenoising dequantization transform
5.761212.098118.43524.7719Oct 31, 2017Dec 23, 2018Feb 14, 2020Apr 8, 2021May 31, 2022Jul 23, 2023Sep 14, 2024

Evaluation results

MethodLinks
denoising dequantization transformPrecision=A4W4, Steps=100k2024.0930.17——
denoising dequantization transformPrecision=A4W2, Steps=100k2024.0930.03——
TransformerPrecision=A32W32, Steps=100k2024.0929.8——
denoising dequantization transformPrecision=A4W4, Steps=25k2024.0929.71——
TransformerPrecision=A32W32, Steps=25k2024.0929.49——
LayerNorm (Ott et al., 2018)Normalization=true2019.0129.3——
Fixup-initNormalization=false2019.0129.3——
denoising dequantization transformPrecision=A2W2, Steps=100k2024.0928.61——
denoising dequantization transformPrecision=A4W2, Steps=25k2024.0928.58——
Vaswani et al. (2017)Normalization=true2019.0128.4——
denoising dequantization transformPrecision=A4W1, Steps=100k2024.0928.32——
denoising dequantization transformPrecision=A2W2, Steps=25k2024.0927.56——
denoising dequantization transformPrecision=A2W1, Steps=100k2024.0927.4——
denoising dequantization transformPrecision=A4W1, Steps=25k2024.0927.06——
denoising dequantization transformPrecision=A2W1, Steps=25k2024.0926——
denoising dequantization transformPrecision=A1W1, Steps=100k2024.0923.64——
denoising dequantization transformPrecision=A1W1, Steps=25k2024.0921.42——
SupervisedDecoding=greedy, Supervision=supervised2017.1021.33——
oracle word reorderingDecoding=greedy, Supervision=unsupervised2017.1011.57——
Our modelDecoding=greedy, Iteration=2nd, Supervision=unsupervised2017.109.75——
Our modelDecoding=greedy, Iteration=3rd, Supervision=unsupervised2017.109.64——
Our modelDecoding=greedy, Iteration=1st, Supervision=unsupervised2017.108.86——
word-by-wordDecoding=greedy, Supervision=unsupervised2017.107.06——
word reorderingDecoding=greedy, Supervision=unsupervised2017.106.7——
Fixed-PointGE=Y, PT=Y, Clip=Y, LB=42024.09—28.4826.94
GradScaleGE=Y, PT=N, Clip=N, LB=42024.09—27.525.9
LSQ+LUQGE=Y, PT=N, Clip=Y, LB=42024.09—27.527.17
LUQ+SMPGE=N, PT=Y, Clip=Y, LB=42024.09—27.527.25
OursGE=N, PT=N, Clip=N, LB=12024.09—29.4929.71