OPT-IML: Scaling Language Model Instruction Meta Learning through the Lens of Generalization

Benchmark Model Rank Results
natural-language-inference-on-rteOPT-IML 175B#26Accuracy: 84.8%
natural-language-inference-on-rteOPT-IML 30B#29Accuracy: 83.8%
natural-language-inference-on-rteOPT-IML 1.3B#59Accuracy: 66.8%
natural-language-inference-on-rteOPT 175B#67Accuracy: 60.3%
natural-language-inference-on-rteOPT 30B#71Accuracy: 58.1%
natural-language-inference-on-rteOPT 1.3B#79Accuracy: 54.2%
question-answering-on-boolqOPT-IML 175B#67Accuracy: 71.4
question-answering-on-boolqOPT-IML 30B#73Accuracy: 66.9
question-answering-on-boolqOPT 30B (0-shot)#81Accuracy: 64
question-answering-on-boolqOPT-IML 1.3B (0-shot)#93Accuracy: 61.5
question-answering-on-boolqOPT 1.3B (zero-shot)#101Accuracy: 60.5
question-answering-on-boolqOPT 175B#104Accuracy: 60.1