List your company

Natural Language Understanding on GLUE v1 (dev)

93.8MRPC ScoreAutoBERT-Zero*
73.41678.7088489.292Jul 12, 2021Nov 1, 2021Feb 21, 2022Jun 14, 2022Oct 4, 2022Jan 24, 2023May 17, 2023

Evaluation results

MethodLinks
AutoBERT-Zero*#Params=104M, Infer FLOPs=2.3e10, initialization=supernet2021.0793.867.386.486.390.885.291.792.595.287.7—
AutoBERT-Zero#Params=104M, Infer FLOPs=2.3e102021.0793.364.585.585.390.881.988.99294.286.3—
AutoBERT-w/o-desc#Params=104M, Infer FLOPs=2.3e10, architecture=backbone without descending kernel sizes2021.0792.865.184.58590.578.788.291.693.785.6—
AutoBERT-conv#Params=104M, Infer FLOPs=2.2e10, architecture=three continuous convolution layers2021.0792.663.884.484.690.180.588.391.793.585.5—
AutoBERT-att#Params=104M, Infer FLOPs=2.3e10, architecture=three continuous attention layers2021.0792.265.484.68590.481.688.591.893.885.9—
BERTbase (Teacher)#Params=110M2023.0591.460.384.984.889.469.791.591.793.784.1—
AD-KD#Params=66M2023.0591.258.383.484.289.270.991.291.291.983.4—
MGSKD#Params=66M2023.0589.849.183.383.988.567.991.290.391.781.5—
BERT-base(ours)#Params=110M, Infer FLOPs=2.9e10, implementation=reproduced by current paper authors2021.0789.758.184.885.288.86988.291.592.983.1—
Flexible Multi-Task Model (w/o KD)Layers=7 + 60, Overhead=67 (69.8%)2021.0789.755.2——88.971.591.191.793.183.284.5
CKD#Params=66M2023.0589.655.183.684.18967.391.290.59382.4—
Vanilla KD#Params=66M2023.0589.453.682.783.188.766.890.590.191.181.6—
PD#Params=66M2023.0589.4—82.583.4—66.790.789.491.1——
BERT6 (Student)#Params=66M2023.0589.251.281.782.688.366.190.489.39180.9—
BERT-of-TheseusLayers=6 x 8, Overhead=48 (50.0%)2021.078951.1——88.768.289.689.591.581.282.3
TinyBERT#Params=66M2023.0588.853.883.183.488.366.990.589.992.381.7—
TinyBERTLayers=6 x 8, Overhead=48 (50.0%)2021.0788.442.8———72.290.690.591.6—83.5
Full fine-tuningLayers=12 x 8, Overhead=96 (100%)2021.0788.254.7——88.869.791.191.693.482.884.6
Flexible Multi-Task Model (KD-3)Layers=7 + 3 x 8, Overhead=31 (32.3%)2021.078850——88.966.891.290.292.781.382.9
Flexible Multi-Task Model (mixed)Layers=7 + 26, Overhead=33 (34.3%)2021.078855.2——88.371.59190.292.782.582.9
DistillBERTLayers=6 x 8, Overhead=48 (50.0%)2021.0787.551.3——86.959.988.589.291.379.682.2
MT-DNN (full)Layers=12 x 1, Overhead=12 (12.5%)2021.0787.451.3——86.880.987.691.193.582.984.4
PD-BERTLayers=6 x 8, Overhead=48 (50.0%)2021.0787.2————66.789.18991.1—83
Flexible Multi-Task Model (KD-2)Layers=7 + 2 x 8, Overhead=23 (24.0%)2021.0786.344——88.664.691.388.692.779.781.7
Vanilla-KDLayers=6 x 8, Overhead=48 (50.0%)2021.0786.245.1——84.964.988.18890.578.580.1
BERT-PKDLayers=6 x 8, Overhead=48 (50.0%)2021.0785.745.5——86.266.588.488.491.379.281.3
MiniLMLayers=6 x 8, Overhead=48 (50.0%)2021.0785.745.5——86.266.588.488.491.379.281.3
PKD#Params=66M2023.0585.745.581.3—86.266.588.488.491.379.2—
Flexible Multi-Task Model (KD-1)Layers=7 + 1 x 8, Overhead=15 (15.6%)2021.0785.136.4——88.366.19186.490.777.477.5
MT-DNN (LOO)2021.0774.210.2——72.960.666.569.779.2—56.7