Charformer: Fast Character Transformers via Gradient-based Subword Tokenization

Benchmark Model Rank Results
linguistic-acceptability-on-colaCharformer-Tall#28Accuracy: 51.8%
natural-language-inference-on-multinliCharformer-Tall#29Matched: 83.7Mismatched: 84.4
natural-language-inference-on-qnliCharformer-Tall#26Accuracy: 91.0%
paraphrase-identification-on-quora-questionCharformer-Tall#3F1: 88.5Accuracy: 91.4
semantic-textual-similarity-on-mrpcCharformer-Tall#21Accuracy: 87.5%F1: 91.4
semantic-textual-similarity-on-sts-benchmarkCharformer-Tall#18Pearson Correlation: 0.873
sentiment-analysis-on-sst-2-binaryCharformer-Base#47Accuracy: 91.6