Read the original document by opening this link in a new tab.
Table of Contents
1 Introduction
2 Scaling law experiments
2.1 Scaling laws
2.2 Downstream metric evaluations
3 Training dataset
4 Evaluation
4.1 Language proficiency exams
4.2 Classification and question answering
4.3 Reasoning
4.4 Coding
4.5 Translation
4.6 Natural language generation
4.7 Memorization
5 Responsible usage
5.1 Inference-time control
5.2 Recommendations for developers
6 Conclusion
7 Authorship, attribution, and acknowledgements
A Detailed results
A.1 Scaling laws
A.2 Instruction tuning
A.3 Multilingual commonsense reasoning
A.4 Coding
A.5 Natural language generation
B Examples of model capabilities
B.1 Multilinguality
B.2 Creative generation
B.3 Coding
C Language proficiency exams
D Responsible AI
D.1 Dataset analysis
D.2 Evaluation approach
D.3 Dialog uses
D.4 Classification uses
D.5 Translation uses
D.6 Question answering uses
D.7 Language modeling
D.8 Measurement quality rubrics
D.9 CrowdWorksheets
D.10 Model Card
Summary
The PaLM 2 Technical Report introduces a new state-of-the-art language model, PaLM 2, that exhibits enhanced multilingual and reasoning capabilities while being more computationally efficient than its predecessor. Through comprehensive evaluations, PaLM 2 demonstrates improved quality on various tasks across different model sizes, with faster and more efficient inference. The report highlights PaLM 2's robust reasoning abilities, stable performance on responsible AI evaluations, and its control over toxicity at inference time without additional overhead. Overall, PaLM 2 achieves state-of-the-art performance across diverse tasks. The document discusses the distinctions within the PaLM 2 family and emphasizes the importance of understanding user-facing products' performance may differ from reported results. The technical report covers various sections such as scaling law experiments, training datasets, evaluations, responsible usage, and detailed results of model capabilities.