LaMDA: Language Models for Dialog Applications

TLDR

Model scaling improves quality but offers limited gains in safety and factual grounding, which involve aligning responses with human values and enabling consultation of external knowledge sources. The study introduces LaMDA and shows that fine‑tuning with annotated data and external knowledge access improves safety and factual grounding. LaMDA is a Transformer‑based model up to 137 B parameters pre‑trained on 1.56 T words, with safety assessed by a human‑values metric and enhanced via a classifier fine‑tuned on crowdworker annotations, and its use in education and recommendation domains was also examined. Fine‑tuning with annotated data and external knowledge access significantly improves safety and factual grounding, as evidenced by higher scores on a human‑values safety metric and a groundedness metric that favors responses linked to known sources.

Abstract

We present LaMDA: Language Models for Dialog Applications. LaMDA is a family of Transformer-based neural language models specialized for dialog, which have up to 137B parameters and are pre-trained on 1.56T words of public dialog data and web text. While model scaling alone can improve quality, it shows less improvements on safety and factual grounding. We demonstrate that fine-tuning with annotated data and enabling the model to consult external knowledge sources can lead to significant improvements towards the two key challenges of safety and factual grounding. The first challenge, safety, involves ensuring that the model's responses are consistent with a set of human values, such as preventing harmful suggestions and unfair bias. We quantify safety using a metric based on an illustrative set of human values, and we find that filtering candidate responses using a LaMDA classifier fine-tuned with a small amount of crowdworker-annotated data offers a promising approach to improving model safety. The second challenge, factual grounding, involves enabling the model to consult external knowledge sources, such as an information retrieval system, a language translator, and a calculator. We quantify factuality using a groundedness metric, and we find that our approach enables the model to generate responses grounded in known sources, rather than responses that merely sound plausible. Finally, we explore the use of LaMDA in the domains of education and content recommendations, and analyze their helpfulness and role consistency.