3 ms·
Large Language Models Can Self-Improve
- zone411 4y agoNote: 540B parameters -> Google Research's PaLM (https://arxiv.org/abs/2204.02311 https://arxiv.org/abs/2204.02311). GSM8K = grade school math word problems DROP = reading comprehension benchmark requiring discrete reasoning over paragraphs OpenBookQA = question-answering dataset modeled after open book exams for assessing human understanding of a subject. 5,957 multiple-choice elementary-level science questions ANLI-A3 = adversarial benchmark designed to be challenging to current state-of-the-art models