Finetuning language models on a collection of datasets phrased as instructions has been shown to improve model performance and generalization to unseen tasks. In this paper we explore instruction finetuning with a particular focus on (1) scaling the number of tasks, (2) scaling the model size, and (3) finetuning on chain-of-thought data. We find that instruction finetuning with the above aspects dramatically improves performance on a variety of model classes (PaLM, T5, U-PaLM), prompting setups (zero-shot, few-shot, CoT), and evaluation benchmarks (MMLU, BBH, TyDiQA, MGSM, open-ended generation, RealToxicityPrompts). For instance, Flan-PaLM 540B instruction-finetuned on 1.8K tasks outperforms PaLM 540B by a large margin (+9.4% on average). Flan-PaLM 540B achieves state-of-the-art performance on several benchmarks (at time of release), such as 75.2% on five-shot MMLU. We also publicly release Flan-T5 checkpoints,1 which achieve strong few-shot performance even compared to much larger models, such as PaLM 62B. Overall, instruction finetuning is a general method for improving the performance and usability of pretrained language models.

Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model (PaLM). We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly efficient training across multiple TPU Pods. We demonstrate continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the finetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A significant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.

Large-scale document retrieval systems often utilize two styles of neural network models which live at two different ends of the joint computation vs. accuracy spectrum.The first style is dual encoder (or two-tower) models, where the query and document representations are computed completely independently and combined with a simple dot product operation.The second style is cross-attention models, where the query and document features are concatenated in the input layer and all computation is based on the joint querydocument representation.Dual encoder models are typically used for retrieval and deep re-ranking, while cross-attention models are typically used for shallow re-ranking.In this paper, we present a lightweight architecture that explores this joint cost vs.accuracy trade-off based on multi-vector attention (MVA).We thoroughly evaluate our method on the MS-MARCO passage retrieval dataset and show how to efficiently trade off retrieval accuracy with joint computation and offline document storage cost.We show that a highly compressed document representation and inexpensive joint computation can be achieved through a combination of learned pooling tokens and aggressive downprojection.Our code and model checkpoints are available on GitHub.

We introduce a novel method of generating synthetic question answering corpora by combining models of question generation and answer extraction, and by filtering the results to ensure roundtrip consistency.By pretraining on the resulting corpora we obtain significant improvements on SQuAD2 (Rajpurkar et al., 2018) and NQ (Kwiatkowski et al., 2019), establishing a new state-of-the-art on the latter.Our synthetic data generation models, for both question generation and answer extraction, can be fully reproduced by finetuning a publicly available BERT model (Devlin et al., 2018) on the extractive subsets of SQuAD2 and NQ.We also describe a more powerful variant that does full sequence-to-sequence pretraining for question generation, obtaining exact match and F1 at less than 0.1% and 0.4% from human performance on SQuAD2.

We present the zero-shot entity linking task, where mentions must be linked to unseen entities without in-domain labeled data.The goal is to enable robust transfer to highly specialized domains, and so no metadata or alias tables are assumed.In this setting, entities are only identified by text descriptions, and models must rely strictly on language understanding to resolve the new entities.First, we show that strong reading comprehension models pre-trained on large unlabeled data can be used to generalize to unseen entities.Second, we propose a simple and effective adaptive pre-training strategy, which we term domainadaptive pre-training (DAP), to address the domain shift problem associated with linking unseen entities in a new domain.We present experiments on a new dataset that we construct for this task and show that DAP improves over strong pre-training baselines, including BERT.

We present the Natural Questions corpus, a question answering data set. Questions consist of real anonymized, aggregated queries issued to the Google search engine. An annotator is presented with a question along with a Wikipedia page from the top 5 search results, and annotates a long answer (typically a paragraph) and a short answer (one or more entities) if present on the page, or marks null if no long/short answer is present. The public release consists of 307,373 training examples with single annotations; 7,830 examples with 5-way annotations for development data; and a further 7,842 examples with 5-way annotated sequestered as test data. We present experiments validating quality of the data. We also describe analysis of 25-way annotations on 302 examples, giving insights into human variability on the annotation task. We introduce robust metrics for the purposes of evaluating question answering systems; demonstrate high human upper bounds on these metrics; and establish baseline results using competitive methods drawn from related literature.

Jiatao Gu, Hany Hassan, Jacob Devlin, Victor O.K. Li. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Attentional sequence-to-sequence models have become the new standard for machine translation, but one challenge of such models is a significant increase in training and decoding cost compared to phrase-based systems.Here, we focus on efficient decoding, with a goal of achieving accuracy close the state-of-the-art in neural machine translation (NMT), while achieving CPU decoding speed/throughput close to that of a phrasal decoder.We approach this problem from two angles: First, we describe several techniques for speeding up an NMT beam search decoder, which obtain a 4.4x speedup over a very efficient baseline decoder without changing the decoder output.Second, we propose a simple but powerful network architecture which uses an RNN (GRU/LSTM) layer at bottom, followed by a series of stacked fully-connected layers applied at every timestep.This architecture achieves similar accuracy to a deep recurrent model, at a small fraction of the training and decoding cost.By combining these techniques, our best system achieves a very competitive accuracy of 38.3 BLEU on WMT English-French NewsTest2014, while decoding at 100 words/sec on single-threaded CPU.We believe this is the best published accuracy/speed trade-off of an NMT system.

The problem of automatically generating a computer program from some specification has been studied since the early days of AI. Recently, two competing approaches for `automatic program learning’ have received significant attention: (1) `neural program synthesis’, where a neural network is conditioned on input/output (I/O) examples and learns to generate a program, and (2) `neural program induction’, where a neural network generates new outputs directly using a latent program representation. Here, for the first time, we directly compare both approaches on a large-scale, real-world learning task and we additionally contrast to rule-based program synthesis, which uses hand-crafted semantics to guide the program generation. Our neural models use a modified attention RNN to allow encoding of variable-sized sets of I/O pairs, which achieve 92\% accuracy on a real-world test set, compared to the 34\% accuracy of the previous best neural synthesis approach. The synthesis model also outperforms a comparable induction model on this task, but we more importantly demonstrate that the strength of each approach is highly dependent on the evaluation metric and end-user application. Finally, we show that we can train our neural models to remain very robust to the type of noise expected in real-world data (e.g., typos), while a highly-engineered rule-based system fails entirely.

Most recently proposed methods for Neural Program induction work under the assumption of having a large set of input/output (I/O) examples for learning any given input-output mapping. This paper aims to address the problem of data and computation efficiency of program induction by leveraging information from related tasks. Specifically, we propose two novel approaches for cross-task knowledge transfer to improve program induction in limited-data scenarios. In our first proposal, portfolio adaptation, a set of induction models is pretrained on a set of related tasks, and the best model is adapted towards the new task using transfer learning. In our second approach, meta program induction, a $k$-shot learning approach is used to make a model generalize to new tasks without additional training. To test the efficacy of our methods, we constructed a new benchmark of programs written in the Karel programming language. Using an extensive experimental evaluation on the Karel benchmark, we demonstrate that our proposals dramatically outperform the baseline induction method that does not use knowledge transfer. We also analyze the relative performance of the two approaches and study conditions in which they perform best. In particular, meta induction outperforms all existing approaches under extreme data sparsity (when a very small number of examples are available), i.e., fewer than ten. As the number of available I/O examples increase (i.e. a thousand or more), portfolio adapted program induction becomes the best approach. For intermediate data sizes, we demonstrate that the combined method of adapted meta program induction has the strongest performance.

Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin, Margaret Mitchell, Xiaodong He, Lucy Vanderwende. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2016.

Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh, Lucy Vanderwende, Michel Galley, Margaret Mitchell. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016.

We present a three-pronged approach to improving Statistical Machine Translation (SMT), building on recent success in the application of neural networks to SMT. First, we propose new features based on neural networks to model various non-local translation phenomena. Second, we augment the architecture of the neural network with tensor layers that capture important higher-order interaction among the network units. Third, we apply multitask learning to estimate the neural network parameters jointly. Each of our proposed methods results in significant improvements that are complementary. The overall improvement is +2.7 and +1.8 BLEU points for Arabic-English and Chinese-English translation over a state-of-the-art system that already includes neural network features.

In the last several years, neural network models have significantly improved accuracy in a number of NLP tasks.However, one serious drawback that has impeded their adoption in production systems is the slow runtime speed of neural network models compared to alternate models, such as maximum entropy classifiers.In Devlin et al. (2014), the authors presented a simple technique for speeding up feed-forward embedding-based neural network models, where the dot product between each word embedding and part of the first hidden layer are pre-computed offline.However, this technique cannot be used for hidden layers beyond the first.In this paper, we explore a neural network architecture where the embedding layer feeds into multiple hidden layers that are placed "next to" one another so that each can be pre-computed independently.On a large scale language modeling task, this architecture achieves a 10x speedup at runtime and a significant reduction in perplexity when compared to a standard multilayer network.

Francis Ferraro, Nasrin Mostafazadeh, Ting-Hao Huang, Lucy Vanderwende, Jacob Devlin, Michel Galley, Margaret Mitchell. Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. 2015.

Jacob Devlin, Rabih Zbib, Zhongqiang Huang, Thomas Lamar, Richard Schwartz, John Makhoul. Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2014.

This paper describes a factored approach to incorporating soft source syntactic constraints into a hierarchical phrase-based translation system.In contrast to traditional approaches that directly introduce syntactic constraints to translation rules by explicitly decorating them with syntactic annotations, which often exacerbate the data sparsity problem and cause other problems, our approach keeps translation rules intact and factorizes the use of syntactic constraints through two separate models: 1) a syntax mismatch model that associates each nonterminal of a translation rule with a distribution of tags that is used to measure the degree of syntactic compatibility of the translation rule on source spans; 2) a syntax-based reordering model that predicts whether a pair of sibling constituents in the constituent parse tree of the source sentence should be reordered or not when translated to the target language.The features produced by both models are used as soft constraints to guide the translation process.Experiments on Chinese-English translation show that the proposed approach significantly improves a strong string-to-dependency translation system on multiple evaluation sets.

When performing handwriting recognition on natural language text, the use of a word-level language model (LM) is known to significantly improve recognition accuracy. The most common type of language model, the n-gram model, decomposes sentences into short, overlapping chunks. In this paper, we propose a new type of language model which we use in addition to the standard n-gram LM. Our new model uses the likelihood score from a statistical machine translation system as a reranking feature. In general terms, we automatically translate each OCR hypothesis into another language, and then create a feature score based on how "difficult" it was to perform the translation. Intuitively, the difficulty of translation correlates with how well-formed the input sentence is. In an Arabic handwriting recognition task, we were able to obtain an 0.4% absolute improvement to word error rate (WER) on top of a powerful 5-gram LM.