PaLM 2 is a state-of-the-art language model that improves upon its predecessor, PaLM, in multilingual capabilities, reasoning, and computational efficiency. It is trained using a mixture of objectives and demonstrates significant improvements in performance across various tasks, including natural language generation, translation, and reasoning. PaLM 2 also exhibits robust reasoning capabilities, as evidenced by its performance on the BIG-Bench and other reasoning tasks. It shows stable performance on responsible AI evaluations and enables inference-time control over toxicity without additional overhead. PaLM 2 is based on the Transformer architecture and incorporates a diverse set of research advances, including compute-optimal scaling, improved dataset mixtures, and architectural and objective improvements. The largest model in the PaLM 2 family, PaLM 2-L, is significantly smaller than the largest PaLM model but uses more training compute. Evaluation results show that PaLM 2 models significantly outperform PaLM on a variety of tasks, including natural language generation, translation, and reasoning. These results suggest that model scaling is not the only way to improve performance; performance can be unlocked by meticulous data selection and efficient architecture/objectives. PaLM 2 demonstrates significant multilingual language, code generation, and reasoning abilities. It performs significantly better than PaLM on real-world advanced language proficiency exams and passes exams in all evaluated languages. PaLM 2 includes control tokens to enable inference-time control over toxicity, modifying only a fraction of pre-training as compared to prior work. It also includes special 'canary' token sequences to enable improved measures of memorization across languages. PaLM 2 is trained on a diverse set of sources, including web documents, books, code, mathematics, and conversational data. It is trained on a dataset that includes a higher percentage of non-English data than previous large language models, which is beneficial for multilingual tasks. PaLM 2 is also trained on parallel data covering hundreds of languages in the form of source and target text pairs. This allows the model to learn each language's nuances. PaLM 2 is evaluated on a variety of tasks, including language proficiency exams, classification and question answering, reasoning, coding, translation, and natural language generation. It demonstrates strong performance on these tasks, with improvements over PaLM in many cases. PaLM 2 is also evaluated for potential harms and biases, including memorization and toxic language. It shows lower average rates of verbatim memorization than PaLM and improved multilingual toxicity classification capabilities. PaLM 2 is evaluated on a variety of benchmarks, including the WMT21 translation sets and the FRMT benchmark for few-shot regional machine translation. It shows improvement over both PaLM and the Google Translate production system according to our primary metric: MQM human evaluations by professional translators. PaLM 2 also shows improvement over PaLM in all locales for regional translation. PaLM 2 is evaluated for potential harms and biases in dialog, generative question answeringPaLM 2 is a state-of-the-art language model that improves upon its predecessor, PaLM, in multilingual capabilities, reasoning, and computational efficiency. It is trained using a mixture of objectives and demonstrates significant improvements in performance across various tasks, including natural language generation, translation, and reasoning. PaLM 2 also exhibits robust reasoning capabilities, as evidenced by its performance on the BIG-Bench and other reasoning tasks. It shows stable performance on responsible AI evaluations and enables inference-time control over toxicity without additional overhead. PaLM 2 is based on the Transformer architecture and incorporates a diverse set of research advances, including compute-optimal scaling, improved dataset mixtures, and architectural and objective improvements. The largest model in the PaLM 2 family, PaLM 2-L, is significantly smaller than the largest PaLM model but uses more training compute. Evaluation results show that PaLM 2 models significantly outperform PaLM on a variety of tasks, including natural language generation, translation, and reasoning. These results suggest that model scaling is not the only way to improve performance; performance can be unlocked by meticulous data selection and efficient architecture/objectives. PaLM 2 demonstrates significant multilingual language, code generation, and reasoning abilities. It performs significantly better than PaLM on real-world advanced language proficiency exams and passes exams in all evaluated languages. PaLM 2 includes control tokens to enable inference-time control over toxicity, modifying only a fraction of pre-training as compared to prior work. It also includes special 'canary' token sequences to enable improved measures of memorization across languages. PaLM 2 is trained on a diverse set of sources, including web documents, books, code, mathematics, and conversational data. It is trained on a dataset that includes a higher percentage of non-English data than previous large language models, which is beneficial for multilingual tasks. PaLM 2 is also trained on parallel data covering hundreds of languages in the form of source and target text pairs. This allows the model to learn each language's nuances. PaLM 2 is evaluated on a variety of tasks, including language proficiency exams, classification and question answering, reasoning, coding, translation, and natural language generation. It demonstrates strong performance on these tasks, with improvements over PaLM in many cases. PaLM 2 is also evaluated for potential harms and biases, including memorization and toxic language. It shows lower average rates of verbatim memorization than PaLM and improved multilingual toxicity classification capabilities. PaLM 2 is evaluated on a variety of benchmarks, including the WMT21 translation sets and the FRMT benchmark for few-shot regional machine translation. It shows improvement over both PaLM and the Google Translate production system according to our primary metric: MQM human evaluations by professional translators. PaLM 2 also shows improvement over PaLM in all locales for regional translation. PaLM 2 is evaluated for potential harms and biases in dialog, generative question answering