LaMDA is a family of Transformer-based neural language models designed for dialog applications, with up to 137B parameters and pre-trained on 1.56T words of public dialog data and web text. While model scaling improves quality, it has limited effects on safety and factual grounding. Fine-tuning with annotated data and enabling external knowledge sources significantly improves safety and factual grounding. Safety involves ensuring responses align with human values, while factual grounding enables the model to consult external sources. LaMDA uses a classifier fine-tuned with crowdworker-annotated data to filter unsafe responses and a groundedness metric to ensure responses are based on known sources. The model can generate responses grounded in known sources rather than just sounding plausible. LaMDA is also explored in education and content recommendation domains, where it shows improved helpfulness and role consistency compared to pre-trained models. The model's performance on quality, safety, and groundedness metrics improves with fine-tuning, and it can adapt to specific roles through pre-conditioning. LaMDA's ability to consult external knowledge sources, such as an information retrieval system, enhances factual grounding. The model's performance on safety and groundedness metrics is comparable to crowdworker levels, while quality metrics show significant improvements. LaMDA's ability to generate grounded responses and use external knowledge sources makes it effective for dialog applications. The model's performance on safety and groundedness metrics is still below human levels, but it shows significant improvements with fine-tuning. LaMDA's ability to adapt to specific roles through pre-conditioning and its use of external knowledge sources make it a promising approach for dialog applications.LaMDA is a family of Transformer-based neural language models designed for dialog applications, with up to 137B parameters and pre-trained on 1.56T words of public dialog data and web text. While model scaling improves quality, it has limited effects on safety and factual grounding. Fine-tuning with annotated data and enabling external knowledge sources significantly improves safety and factual grounding. Safety involves ensuring responses align with human values, while factual grounding enables the model to consult external sources. LaMDA uses a classifier fine-tuned with crowdworker-annotated data to filter unsafe responses and a groundedness metric to ensure responses are based on known sources. The model can generate responses grounded in known sources rather than just sounding plausible. LaMDA is also explored in education and content recommendation domains, where it shows improved helpfulness and role consistency compared to pre-trained models. The model's performance on quality, safety, and groundedness metrics improves with fine-tuning, and it can adapt to specific roles through pre-conditioning. LaMDA's ability to consult external knowledge sources, such as an information retrieval system, enhances factual grounding. The model's performance on safety and groundedness metrics is comparable to crowdworker levels, while quality metrics show significant improvements. LaMDA's ability to generate grounded responses and use external knowledge sources makes it effective for dialog applications. The model's performance on safety and groundedness metrics is still below human levels, but it shows significant improvements with fine-tuning. LaMDA's ability to adapt to specific roles through pre-conditioning and its use of external knowledge sources make it a promising approach for dialog applications.