First Train to Generate, then Generate to Train: UnitedSynT5 for Few-Shot NLI

Benchmark Model Rank Results
natural-language-inference-on-e-snliUnitedSynT5 (335M)–Accuracy: 89.8
natural-language-inference-on-e-snliUnitedSynT5 (3B)–Accuracy: 94.0
natural-language-inference-on-multinliUnitedSynT5 (335M)–Matched: 89.8
natural-language-inference-on-multinliUnitedSynT5 (3B)–Matched: 92.6
natural-language-inference-on-snliUnitedSynT5 (335M)–% Test Accuracy: 93.5
natural-language-inference-on-snliUnitedSynT5 (3B)–% Test Accuracy: 94.7