Source-linked AI summary
LaMDA: Language Models for Dialog Applications
Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, Quoc Le
TL;DR
LaMDA addresses the limited safety and factual grounding gains of scaling dialog models alone. It combines large-scale Transformer pre-training with crowd-annotated fine-tuning and external tools. The resulting approach significantly improves the main metrics and application helpfulness, although factuality, complex reasoning, and safety reliability remain limited.
Problem
Scaling improves dialog quality but provides limited gains on safety and groundedness, while plausible model responses can contradict authoritative facts.
Method
LaMDA combines Transformer pre-training with crowdworker-annotated fine-tuning, safety filtering, and external information-retrieval tools for grounded responses.
Results
Fine-tuning significantly improves quality, safety, and groundedness, while LaMDA-based applications are significantly more helpful than pre-training-only applications.
Takeaways & Limitations
Annotated fine-tuning and external API access provide practical paths toward safer, more grounded, and more helpful dialog systems.
Takeaways & Limitations
Groundedness remains limited to simple factual questions, and fine-tuning does not guarantee accurate source reproduction or complete safety reliability.
Abstract
from arXiv · showhide
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.
1 Introduction
LaMDA is a scaled Transformer family for dialog, combining pre-training with safety filtering, external grounding, and response reranking. Scaling improves quality, while fine-tuning produces broader gains across quality, safety, groundedness, and application helpfulness.
- LaMDA models range from 2B to 137B parameters and use 1.56T words of public dialog data and web documents.
- A single LaMDA model generates candidate responses that are safety-filtered, externally grounded, and reranked for quality.
- Scaling alone improves quality, but safety and groundedness remain far below human performance; scaling plus fine-tuning significantly improves all three metrics.
- Safety fine-tuning uses diverse crowdworker labels and a discriminator to detect and remove unsafe candidate responses.
- Groundedness fine-tuning trains LaMDA to use external tools and produce responses supported by known sources.
- Fine-tuned LaMDA models are significantly more helpful in education and content-recommendation applications, while both model types adapt fairly well to assigned roles.
2 Related work
Related work frames LaMDA within research on scaling, dialog modeling, safety, groundedness, and multifaceted evaluation. The paper combines these strands through fine-tuning, external retrieval tools, and separate human-centered metrics.
- Prior work links larger language and dialog models with improved capabilities and dialog quality, motivating LaMDA’s scaling study.
- Retrieval-augmented approaches generally target open-domain question answering, whereas LaMDA trains a dialog model to use an intermediate external tool.
- Automated dialog metrics may correlate poorly with human judgment, so related work motivates evaluating separate, debuggable quality, safety, and groundedness dimensions.
- Existing dialog research identifies risks including toxicity, bias, hate speech, and personally identifying information in training data and model outputs.
- Meaningfully addressing language- and dialog-model risks remains an active research area despite proposed mitigation strategies.
- Groundedness research commonly evaluates whether outputs agree with authoritative sources, including through the two-stage AIS framework.
3 LaMDA pre-training
LaMDA is pre-trained as a general next-token language model on a large mixture of public dialog and web text. The pre-training setup supports later scaling experiments and distinguishes the pre-training-only model as PT.
- LaMDA predicts the next token on public dialog data and web documents, allowing it to serve as a general language-model prior before fine-tuning.
- The dataset contains 2.97B documents, 1.12B dialogs, 13.39B utterances, and 1.56T words, with over 90% in English.
- The largest model has 137B non-embedding parameters and uses a decoder-only Transformer architecture.
- Training used 1024 TPU-v3 chips for about 57.7 days, while additional 2B- and 8B-parameter models measured scaling effects.
- The pre-training stage is summarized in Figure 2, and the model before fine-tuning is called PT, meaning PreTrained.
- PT decoding samples 16 independent top-k candidates with k = 40 and selects the highest-scoring response using log-likelihood and length.
4 Metrics
LaMDA is evaluated with human-centered foundation metrics for quality, safety, and groundedness, alongside application-specific helpfulness and role consistency. The metrics separate conversational quality from safety, factual support, and role performance.
- Foundation metrics: Quality, Safety and Groundedness: Overall quality is the average of sensibleness, specificity, and interestingness, collectively called SSI.
- Foundation metrics: Quality, Safety and Groundedness: Sensibleness measures contextual coherence, while specificity measures whether a response is tailored to the given context.
- Foundation metrics: Quality, Safety and Groundedness: Interestingness is a crowdworker 0/1 judgment that captures attention, curiosity, unexpectedness, wit, or insight.
- Foundation metrics: Quality, Safety and Groundedness: Safety measures unsafe output using objectives derived from Google’s AI Principles, including avoiding harm and unfair bias.
- Foundation metrics: Quality, Safety and Groundedness: Groundedness measures the share of externally informative responses whose claims can be supported by authoritative sources.
- Application metrics: Helpfulness and role consistency evaluate application-specific performance: correctness and usefulness for users, and conformity with the assigned agent role.
5 LaMDA fine-tuning and evaluation data
LaMDA’s evaluation data combines crowdworker judgments for quality, safety, and groundedness, using task-specific dialogs and holdout datasets. The groundedness procedure distinguishes externally verifiable claims from common knowledge and assesses whether claims can be supported by known sources.
- Quality: 6400 dialogs with 121K turns support quality evaluation, where crowdworkers label responses as sensible, specific, and interesting.The Mini-Turing Benchmark adds 1477 dialogs with up to three turns for quality assessment.
- Safety: 8K dialogs with 48K turns support safety evaluation across natural, sensitive-topic, and adversarial interactions.A holdout dataset contains 1166 dialogs with 1458 turns, and responses receive safety scores from three crowdworkers.
- Groundedness: 4K dialogs with 40K turns support groundedness evaluation through crowdworker interactions focused on information-seeking.Workers assess whether turns make external-world claims and whether those claims are known to be true.
- Groundedness: Groundedness data excludes claims about publicly unrecognized personas and treats claims known by three crowdworkers as common knowledge.Claims about historical people, unlike improvised personal claims, require checking against external sources.
- Groundedness: 784 dialog turns from Dinan et al. provide a groundedness evaluation set, with three crowdworkers rating factual claims and their verifiability against known sources.Human-generated responses are also sampled and produced with instructions to use external tools when necessary.
6 LaMDA fine-tuning
LaMDA uses fine-tuning to make one model both generate responses and discriminate among them for quality and safety, while also learning to consult external tools for grounding. Candidate responses are filtered for safety, ranked for quality, and iteratively checked against retrieved information before a final response is produced.
- 6 LaMDA fine-tuning: LaMDA combines generative and discriminative fine-tuning so one model can generate responses and evaluate their quality and safety.Generative examples predict responses, while discriminative examples predict attribute ratings.
- 6.1 Discriminative and generative fine-tuning for Quality (SSI) and Safety: Safety filtering precedes quality ranking: candidates below a safety threshold are removed, and remaining responses are ranked using SSI predictions.Sensibleness receives three times the weight of specificity and interestingness in the ranking procedure.
- 6.1 Discriminative and generative fine-tuning for Quality (SSI) and Safety: 2.5M sampled dialog turns are filtered into 800K safe, sensible, specific, and interesting turns for subsequent response-generation fine-tuning.The paper reports significant gains in safety and quality from this technique.
- 6.2 Fine-tuning to learn to call an external information retrieval system: External grounding addresses plausible responses that contradict established facts, although grounding in known sources does not guarantee factual accuracy.The approach is intended to let users or external systems judge response validity.
- 6.2 Fine-tuning to learn to call an external information retrieval system: The toolset combines information retrieval, calculation, and translation, while LaMDA learns to generate queries and incorporate returned snippets.The toolset accepts a string and returns one or more strings, such as an age lookup, arithmetic result, or translation.
- 6.2 Fine-tuning to learn to call an external information retrieval system: The groundedness pipeline uses a Base draft followed by Research calls that either query the toolset or address the user.The first generated token directs each step to the toolset or user, with a maximum of four research iterations in the example.
- 6.2 Fine-tuning to learn to call an external information retrieval system: The Research phase replaces an unverified claim with a source-based statement when retrieved results provide no evidence for the original claim.In the example, the final response preserves a similar dialog tone while changing the unsupported content.
7 Results on foundation metrics
Across model sizes, fine-tuning substantially improves LaMDA’s quality, safety, and groundedness, while scaling alone provides limited safety gains. Groundedness fine-tuning enables external retrieval, reaching 73.2% groundedness and 65% citation accuracy.
- Fine-tuning the pre-trained model trains quality and safety discriminators that filter unsafe candidates and re-rank responses by sensibleness, specificity, and interestingness.
- LaMDA fine-tunes the quality-safety model to call an external information retrieval system and provide attributed responses.
- Fine-tuning significantly improves quality, safety, and groundedness across all model sizes, while quality generally also improves with model size.
- Scaling without fine-tuning provides little safety benefit, whereas combining scaling with safety fine-tuning significantly improves safety.
- 73.2% Groundedness and 65% Citation Accuracy are achieved by shifting some knowledge-retention demands to external sources.Groundedness measures responses attributable to known sources; citation accuracy measures whether required responses include source URLs.
- Fine-tuning reaches 92.3% sensibleness, potentially requiring a dense model multiple orders of magnitude larger than the 137B-parameter pre-trained model.
8 Domain grounding
LaMDA uses brief role-specific dialog preconditioning to ground agents in educational and music-recommendation domains. In evaluations, LaMDA applications were more helpful than pre-training-only applications, while role consistency was generally high but factual and recommendation errors remained.
- Domain grounding: LaMDA explores domain grounding by acting as Mount Everest for education and as a music recommendation agent.Each role is specified with a brief domain description.
- Domain grounding: A single role-aligned greeting preconditions the Mount Everest agent, while Music receives longer preconditioning to establish role and interaction style.The same role-specific preconditioning is applied to LaMDA and pre-training-only models.
- Evaluation: Evaluation uses 600 dialog turns, with each response labeled three times for helpfulness and role consistency by crowdworkers given the role definitions.Crowdworkers first interact with the agents and then label responses in their original context.
- Results: LaMDA applications perform significantly better than pre-training-only applications on helpfulness, while both types generally score well on role consistency.Role consistency occasionally breaks, including Mount Everest referring to itself in the third person.
- Results: Grounding errors reduce helpfulness: Mount Everest facts lack known-source attribution in about 30% of responses, while Music misses recommendations in about 9% and gives broken links in about 7%.Crowdworkers verify links and information against known sources during evaluation.
9 Discussion and limitations
LaMDA shows that modest human-annotated fine-tuning can improve dialog quality and safety, but substantial limitations remain in factual grounding, bias, evaluation, and deployment scope.
- Discussion: Less than 0.001% of pre-training data enabled significant progress toward better quality and safer dialog models.The authors nevertheless characterize LaMDA and the study as having many limitations.
- Evaluation: Crowdworker-based evaluation is constrained by expensive data collection, subjective judgments, limited disagreement analysis, and a population that may not represent users.Future work includes broader recruitment, better label quality, and methods accounting for systematic social and cultural disagreement.
- Limitations: Fine-tuning can improve groundedness, but responses may still misrepresent authoritative sources, especially beyond simple factual questions.More complex reasoning remains open, alongside subtler conversational quality failures.
- Safety: Safety fine-tuning improves average safety metrics but does not guarantee reliability against long-tail inappropriate responses or capture all dimensions of safety and fairness.The authors call for more robust standards for general-purpose dialog models.
- Deployment scope: LaMDA is presented as a recipe for eventually producing production-ready systems for specific applications, not as a final model.Real-world deployment still faces challenges including training-data bias, representational harms, deception, manipulation, and impersonation risks.
- Bias and values: Safety objectives and rating scales cannot universally encode cultural norms, social appropriateness, or nuanced differences in harmfulness.The paper advocates context-sensitive and participatory approaches involving relevant communities.
- Future work: Further research is expected to yield additional performance gains because current findings are preliminary and rely on relatively modest fine-tuning data.The authors remain encouraged despite the limitations of the current approach.
10 Energy and Carbon Footprint Estimate of LaMDA
The estimate reports the computational resources and energy-related parameters used to pre-train LaMDA’s largest model.
- Compute estimate: 3.55E+23 total FLOPS were used to pre-train LaMDA’s largest model, exceeding GPT-3’s reported 3.14E+23 FLOPS.Training used 1024 TPU-V3 chips for 57.7 days at 56.5% FLOPS utilization and 123 TFLOPS/s.
- Energy estimate: The datacenter PUE was 1.10, and measured average power per TPUv3 accelerator was roughly 289W.The power estimate borrowed measurements from LaMDA-related Meena experiments.
11 Conclusion
The paper finds that scaling improves dialog metrics, while crowd-annotated fine-tuning and external APIs provide additional gains in safety, quality, and groundedness. It also reports that LaMDA-based applications are more helpful than pre-training-only applications, while safety and groundedness remain below human performance.
- Scaling alone improves all metrics, but its safety and groundedness gains remain far behind human performance.
- Crowd-annotated data drives significant additional gains in the evaluated metrics.
- LaMDA-based applications are significantly more helpful than pre-training-only applications under the same application-specific preconditioning.
- The paper frames LaMDA as a step closer to practical and safe open-ended dialog systems.
- The safety objectives are illustrative rather than exhaustive, and the crowdworker population may not fully reflect the user base.
A.4 Safety annotations data distribution
The appendix documents how crowdworkers evaluated dialog quality and groundedness and presents selected model-response examples. The instructions require judgments about sensibleness, specificity, interestingness, factual claims, and factual correctness.
- SSI evaluation asks whether a response makes sense, is specific to context, and is interesting.
- Groundedness evaluation first checks whether the conversation is understandable and whether the response contains factual claims.
- Factual correctness is assessed by verifying every factual claim with an information retrieval tool and reliable sources.
- Selected tables compare pre-trained and fine-tuned responses across adversarial, prompted, and Wizard-of-Wikipedia contexts.
- The examples show that external tool usage can raise groundedness and encourage citations, while unsupported factual claims remain possible.
- The examples also document errors in reasoning, programming, factual completeness, formatting, and theory-of-mind questions.
E Pre-training data composition
LaMDA’s Infiniset pre-training data combines public dialog data with several categories of public web documents.
- Infiniset contains 2.97B documents, 1.12B dialogs, and 13.39B utterances.
- Half of the corpus consists of dialog data from public forums.
- The remaining data includes C4, code documents, English Wikipedia, English web documents, and non-English web documents.
F Pre-training and fine-tuning results
The section presents results for the paper’s foundation metrics.
- Table 28 reports results for the foundation metrics.
- The supplied passage identifies the results table but provides no metric values.
- The table is presented as part of the pre-training and fine-tuning results section.