Research papers by Cohere For AI and Technical Staff at Cohere

Research papers | Cohere For AI

C4AI - Research Papers

As Large Language Models (LLMs) have become more advanced, they have outpaced our abilities to accurately evaluate their quality. Not only is finding data to adequately probe particular model properties difficult, but evaluating the correctness of a model's freeform generation alone is a challenge. To address this, many evaluations now rely on using LLMs themselves as judges to score the quality of outputs from other LLMs. Evaluations most commonly use a single large model like GPT4. While this method has grown in popularity, it is costly, has been shown to introduce intramodel bias, and in this work, we find that very large models are often unnecessary. We propose instead to evaluate models using a Panel of LLm evaluators (PoLL). Across three distinct judge settings and spanning six different datasets, we find that using a PoLL composed of a larger number of smaller models outperforms a single large judge, exhibits less intra-model bias due to its composition of disjoint model families, and does so while being over seven times less expensive.

Human feedback plays a central role in the alignment of Large Language Models (LLMs). However, open questions remain about the methods (how), domains (where), people (who) and objectives (to what end) of human feedback collection. To navigate these questions, we introduce PRISM, a new dataset which maps the sociodemographics and stated preferences of 1,500 diverse participants from 75 countries, to their contextual preferences and fine-grained feedback in 8,011 live conversations with 21 LLMs. PRISM contributes (i) wide geographic and demographic participation in human feedback data; (ii) two census-representative samples for understanding collective welfare (UK and US); and (iii) individualised feedback where every rating is linked to a detailed participant profile, thus permitting exploration of personalisation and attribution of sample artefacts. We focus on collecting conversations that centre subjective and multicultural perspectives on value-laden and controversial topics, where we expect the most interpersonal and cross-cultural disagreement. We demonstrate the usefulness of PRISM via three case studies of dialogue diversity, preference diversity, and welfare outcomes, showing that it matters which humans set alignment norms. As well as offering a rich community resource, we advocate for broader participation in AI development and a more inclusive approach to technology design.

To date, toxicity mitigation in language models has almost entirely been focused on single-language settings. As language models embrace multilingual capabilities, it’s crucial our safety measures keep pace. Recognizing this research gap, our approach expands the scope of conventional toxicity mitigation to address the complexities presented by multiple languages. In the absence of sufficient annotated datasets across languages, we employ translated data to evaluate and enhance our mitigation techniques. We also compare finetuning mitigation approaches against retrieval-augmented techniques under both static and continual toxicity mitigation scenarios. This allows us to examine the effects of translation quality and the cross-lingual transfer on toxicity mitigation. We also explore how model size and data quantity affect the success of these mitigation efforts. Covering nine languages, our study represents a broad array of linguistic families and levels of resource availability, ranging from high to mid-resource languages. Through comprehensive experiments, we provide insights into the complexities of multilingual toxicity mitigation, offering valuable insights and paving the way for future research in this increasingly important field. Code and data are available at https://github.com/for-ai/goodtriever.

Work to date on toxicity mitigation has exclusively focused on the English language. Yet harm exists in all languages. We delve into the complexities of toxicity assessment across 9 languages spanning 5 scripts, ranging from high to mid-resource availability. 

The democratization of pre-trained language models through open-source initiatives has rapidly advanced innovation and expanded access to cutting-edge technologies. However, this openness also brings significant security risks, including backdoor attacks, where hidden malicious behaviors are triggered by specific inputs, compromising natural language processing (NLP) system integrity and reliability. This paper suggests that merging a backdoored model with other homogeneous models can remediate backdoor vulnerabilities even if such models are not entirely secure. In our experiments, we explore various models (BERT-Base, RoBERTa-Large, Llama2-7B, and Mistral-7B) and datasets (SST-2, OLID, AG News, and QNLI). Compared to multiple advanced defensive approaches, our method offers an effective and efficient inference-stage defense against backdoor attacks without additional resources or specific knowledge. Our approach consistently outperforms the other advanced baselines, leading to an average of 75% reduction in the attack success rate. Since model merging has been an established approach for improving model performance, the extra advantage it provides regarding defense can be seen as a cost-free bonus.

This paper studies the evolving domain of Continual Learning (CL) in large language models (LLMs), with a focus on developing strategies for efficient and sustainable training. Our primary emphasis is on continual domain-adaptive pretraining, a process designed to equip LLMs with the ability to integrate new information from various domains while retaining previously learned knowledge and enhancing cross-domain knowledge transfer without relying on domain-specific identification. Unlike previous studies, which mostly concentrate on a limited selection of tasks or domains and primarily aim to address the issue of forgetting, our research evaluates the adaptability and capabilities of LLMs to changing data landscapes in practical scenarios. To this end, we introduce a new benchmark designed to measure the adaptability of LLMs to these evolving data environments, offering a comprehensive framework for evaluation. We examine the impact of model size on learning efficacy and forgetting, as well as how the progression and similarity of emerging domains affect the knowledge transfer within these models. Our findings uncover several key insights: (i) when the sequence of domains shows semantic similarity, continual pretraining enables LLMs to better specialize in the current domain compared to stand-alone fine-tuning, (ii) training across a diverse range of domains enhances both backward and forward knowledge transfer, and (iii) smaller models are particularly sensitive to continual pretraining, showing the most significant rates of both forgetting and learning. We posit that our research marks a shift towards establishing a more realistic benchmark for investigating CL in LLMs, and has the potential to play a key role in guiding the direction of future research in the field.

Curious about how language models manage to keep learning new things over time? We investigated continual pretraining in large language models on a diverse set of domains - our findings reveal new insights about knowledge transfer and forgetting.

AI alignment in the shape of Reinforcement Learning from Human Feedback (RLHF) is increasingly treated as a crucial ingredient for high performance large language models. \textsc{Proximal Policy Optimization} (PPO) has been positioned by recent literature as the canonical method for the RL part of RLHF. However, it involves both high computational cost and sensitive hyperparameter tuning. We posit that most of the motivational principles that led to the development of PPO are less of a practical concern in RLHF and advocate for a less computationally expensive method that preserves and even increases performance. We revisit the \textit{formulation} of alignment from human preferences in the context of RL. Keeping simplicity as a guiding principle, we show that many components of PPO are unnecessary in an RLHF context and that far simpler REINFORCE-style optimization variants outperform both PPO and newly proposed "RL-free" methods such as DPO and RAFT. Our work suggests that careful adaptation to LLMs alignment characteristics enables benefiting from online RL optimization at low cost.

Reinforcement learning from human feedback has been widely adopted as a way to ensure models reflect preferences. Approaches like PPO directly borrow from traditional RL assumptions. Is this necessary for LLM settings? 

Datasets are foundational to many breakthroughs in modern artificial intelligence. Many recent achievements in the space of natural language processing (NLP) can be attributed to the finetuning of pre-trained models on a diverse set of tasks that enables a large language model (LLM) to respond to instructions. Instruction fine-tuning (IFT) requires specifically constructed and annotated datasets. However, existing datasets are almost all in the English language. In this work, our primary goal is to bridge the language gap by building a human-curated instruction-following dataset spanning 65 languages. We worked with fluent speakers of languages from around the world to collect natural instances of instructions and completions. Furthermore, we create the most extensive multilingual collection to date, comprising 513 million instances through templating and translating existing datasets across 114 languages. In total, we contribute four key resources: we develop and open-source the Aya Annotation Platform, the Aya Dataset, the Aya Collection, and the Aya Evaluation Suite. The Aya initiative also serves as a valuable case study in participatory research, involving collaborators from 119 countries. We see this as a valuable framework for future research collaborations that aim to bridge gaps in resources.

The Aya Collection stands as the most extensive assembly of multilingual instruction fine-tuning datasets to date, featuring 513 million prompts and completions across 114 languages. We fully open-source the collection, which includes rare, human-curated annotations from fluent speakers worldwide.

Recent breakthroughs in large language models (LLMs) have centered around a handful of data-rich languages. What does it take to broaden access to breakthroughs beyond first-class citizen languages? Our work introduces Aya, a massively multilingual generative language model that follows instructions in 101 languages of which over 50% are considered as lower-resourced. Aya outperforms mT0 and BLOOMZ on the majority of tasks while covering double the number of languages. We introduce extensive new evaluation suites that broaden the state-of-art for multilingual eval across 99 languages -- including discriminative and generative tasks, human evaluation, and simulated win rates that cover both held-out tasks and in-distribution performance. Furthermore, we conduct detailed investigations on the optimal finetuning mixture composition, data pruning, as well as the toxicity, bias, and safety of our models. We open-source our instruction datasets and our model at https://hf.co/CohereForAI/aya-101

The Aya Model is a massively multilingual language model capable of following instructions in 101 languages. Developed using a diverse mix of instructions from the Aya dataset and collection among others, it achieves state-of-the-art performance across numerous multilingual benchmarks.

In the machine learning ecosystem, hardware selection is often regarded as a mere utility, overshadowed by the spotlight on algorithms and data. This oversight is particularly problematic in contexts like ML-as-a-service platforms, where users often lack control over the hardware used for model deployment. How does the choice of hardware impact generalization properties? This paper investigates the influence of hardware on the delicate balance between model performance and fairness. We demonstrate that hardware choices can exacerbate existing disparities, attributing these discrepancies to variations in gradient flows and loss surfaces across different demographic groups. Through both theoretical and empirical analysis, the paper not only identifies the underlying factors but also proposes an effective strategy for mitigating hardware-induced performance imbalances.

How does the choice of hardware impact generalization properties? This paper investigates the influence of hardware on the delicate balance between model performance and fairness.

In Natural Language Processing (NLP), the Elo rating system, originally designed for ranking players in dynamic games such as chess, is increasingly being used to evaluate Large Language Models (LLMs) through "A vs B" paired comparisons. However, while popular, the system's suitability for assessing entities with constant skill levels, such as LLMs, remains relatively unexplored. We study two fundamental axioms that evaluation methods should adhere to: reliability and transitivity. We conduct extensive evaluation of Elo behaviour, illustrating that individual Elo computations exhibit volatility and delving into the impact of varying the Elo rating system's hyperparameters. We show that these axioms are not always satisfied raising questions about the reliability of current comparative evaluations of LLMs. If the current use of Elo scores is intended to substitute the costly head-to-head comparison of LLMs, it is crucial to ensure the ranking is as robust as possible. Guided by the axioms, our findings offer concrete guidelines for enhancing the reliability of LLM evaluation methods, suggesting a need for reassessment of existing comparative approaches.

Elo has proven effective for dynamic games like chess and has recently seen widespread use for evaluating LLMs. But how reliable is it for evaluating static-skill entities like LLMs? We find scenarios where it’s not!


The race to train language models on vast, diverse, and inconsistently documented datasets has raised pressing concerns about the legal and ethical risks for practitioners. To remedy these practices, threatening data transparency and understanding, we convene a multi-disciplinary effort between legal and machine learning experts to systematically audit and trace 1800+ finetuning datasets. Our landscape analysis highlights the sharp divides in composition and focus of commercially open vs closed datasets, with closed datasets monopolizing important categories: lower resource languages, more creative tasks, richer topic variety, newer and more synthetic training data. This points to a deepening divide in the types of data that are made available under different license conditions, and heightened implications for jurisdictional legal interpretations of copyright and fair use. We also observe frequent miscategorization of licenses on widely used dataset hosting sites, with license omission of 72%+ and error rates of 50%+. This points to a crisis in misattribution, and informed use of the most popular datasets, driving many recent breakthroughs. As a contribution to ongoing improvements in dataset transparency and responsible use, we release our entire audit, with an interactive UI, the Data Provenance Explorer, which allows practitioners to trace and filter on data provenance for the most popular open source finetuning data collections: www.dataprovenance.org.

A cross-institutional effort involving experts across 13 institutions to shine a spotlight on data transparency and attribution in AI. dataprovenance.org

Numerous studies have highlighted the privacy risks associated with pretrained large language models. In contrast, our research offers a unique perspective by demonstrating that pretrained large language models can effectively contribute to privacy preservation. We propose a locally differentially private mechanism called DP-Prompt, which leverages the power of pretrained large language models and zero-shot prompting to counter author de-anonymization attacks while minimizing the impact on downstream utility. When DP-Prompt is used with a powerful language model like ChatGPT (gpt-3.5), we observe a notable reduction in the success rate of deanonymization attacks, showing that it surpasses existing approaches by a considerable margin despite its simpler design. For instance, in the case of the IMDB dataset, DP-Prompt (with ChatGPT) perfectly recovers the clean sentiment F1 score while achieving a 46% reduction in author identification F1 score against static attackers and a 26% reduction against adaptive attackers. We conduct extensive experiments across six open-source large language models, ranging up to 7 billion parameters, to analyze various effects of the privacy-utility tradeoff.

Human evaluation is increasingly critical for assessing large language models, capturing linguistic nuances, and reflecting user preferences more accurately than traditional automated metrics. However, the resource-intensive nature of this type of annotation process poses significant challenges. The key question driving our work: "is it feasible to minimize human-in-the-loop feedback by prioritizing data instances which most effectively distinguish between models?" We evaluate several metric-based methods and find that these metrics enhance the efficiency of human evaluations by minimizing the number of required annotations, thus saving time and cost, while ensuring a robust performance evaluation. We show that our method is effective across widely used model families, reducing instances of indecisive (or "tie") outcomes by up to 54% compared to a random sample when focusing on the top-20 percentile of prioritized instances. This potential reduction in required human effort positions our approach as a valuable strategy in future large language model evaluations.

Human evaluation of LLMs is critical, but comes at a high cost.

This research proposes prompt ranking methods to make pairwise human evaluation more efficient. 

Considerable effort has been dedicated to mitigating toxicity, but existing methods often require drastic modifications to model parameters or the use of computationally intensive auxiliary models. Furthermore, previous approaches have often neglected the crucial factor of language's evolving nature over time. In this work, we present a comprehensive perspective on toxicity mitigation that takes into account its changing nature. We introduce Goodtriever, a flexible methodology that matches the current state-of-the-art toxicity mitigation while achieving 43% relative latency reduction during inference and being more computationally efficient. By incorporating a retrieval-based approach at decoding time, Goodtriever enables toxicity-controlled text generation. Our research advocates for an increased focus on adaptable mitigation techniques, which better reflect the data drift models face when deployed in the wild

Toxicity definitions evolve and change over time. Why don't mitigation techniques account for this?

With Goodtriever, we take the first steps on continual toxicity mitigation!



Pushing the boundaries of machine learning often requires exploring different hardware and software combinations. However, the freedom to experiment across different tooling stacks can be at odds with the drive for efficiency, which has produced increasingly specialized AI hardware and incentivized consolidation around a narrow set of ML frameworks. Exploratory research can be restricted if software and hardware are co-evolving, making it even harder to stray away from mainstream ideas that work well with popular tooling stacks. While this friction increasingly impacts the rate of innovation in machine learning, to our knowledge the lack of portability in tooling has not been quantified. In this work, we ask: How portable are popular ML software frameworks? We conduct a large-scale study of the portability of mainstream ML frameworks across different hardware types. Our findings paint an uncomfortable picture – frameworks can lose more than 40% of their key functions when ported to other hardware. Worse, even when functions are portable, the slowdown in their performance can be extreme and render performance untenable. Collectively, our results reveal how costly straying from a narrow set of hardware-software combinations can be - and suggest that specialization of hardware impedes innovation in machine learning research.

How portable are popular ML software frameworks? This work reveals how costly straying from a narrow set of hardware-software combinations can be.

The Mixture of Experts (MoE) is a widely known neural architecture where an ensemble of specialized sub-models optimizes overall performance with a constant computational cost. However, conventional MoEs pose challenges at scale due to the need to store all experts in memory. In this paper, we push MoE to the limit. We propose extremely parameter-efficient MoE by uniquely combining MoE architecture with lightweight experts. Our MoE architecture outperforms standard parameter-efficient fine-tuning (PEFT) methods and is on par with full fine-tuning by only updating the lightweight experts -- less than 1% of an 11B parameters model. Furthermore, our method generalizes to unseen tasks as it does not depend on any prior task knowledge. Our research underscores the versatility of the mixture of experts architecture, showcasing its ability to deliver robust performance even when subjected to rigorous parameter constraints. Our code used in all the experiments is publicly available here: 

Can we use Mixture of Experts (MoE) for instruction tuning in extreme parameter constraint? Here, we push MoEs to the limit with ultra-lightweight experts, enabling parameter-efficient MoEs with high performance.

Large volumes of text data have contributed significantly to the development of large language models (LLMs) in recent years. This data is typically acquired by scraping the internet, leading to pretraining datasets comprised of noisy web text. To date, efforts to prune these datasets down to a higher quality subset have relied on hand-crafted heuristics encoded as rule-based filters. In this work, we take a wider view and explore scalable estimates of data quality that can be used to systematically measure the quality of pretraining data. We perform a rigorous comparison at scale of the simple data quality estimator of perplexity, as well as more sophisticated and computationally intensive estimates of the Error L2-Norm and memorization. These metrics are used to rank and prune pretraining corpora, and we subsequently compare LLMs trained on these pruned datasets. Surprisingly, we find that the simple technique of perplexity outperforms our more computationally expensive scoring methods. We improve over our no-pruning baseline while training on as little as 30% of the original training dataset. Our work sets the foundation for unexplored strategies in automatically curating high quality corpora and suggests the majority of pretraining data can be removed while retaining performance.

Leveraging data pruning to examine what makes “good data.” We explore several metrics for measuring LLM pretraining data and find that we can remove up to 70% of pre-training data while achieving better test set performance.

Large language models exhibit enhanced zero-shot performance on various tasks when fine-tuned with instruction-following data. Multimodal instruction-following models extend these capabilities by integrating both text and images. However, existing models such as MiniGPT-4 face challenges in maintaining dialogue coherence in scenarios involving multiple images. A primary reason is the lack of a specialized dataset for this critical application. To bridge these gaps, we present SparklesChat, a multimodal instruction-following model for open-ended dialogues across multiple images. To support the training, we introduce SparklesDialogue, the first machine-generated dialogue dataset tailored for word-level interleaved multi-image and text interactions. Furthermore, we construct SparklesEval, a GPT-assisted benchmark for quantitatively assessing a model's conversational competence across multiple images and dialogue turns. Our experiments validate the effectiveness of SparklesChat in understanding and reasoning across multiple images and dialogue turns. Specifically, SparklesChat outperformed MiniGPT-4 on established vision-and-language benchmarks, including the BISON binary image selection task and the NLVR2 visual reasoning task. Moreover, SparklesChat scored 8.56 out of 10 on SparklesEval, substantially exceeding MiniGPT-4's score of 3.91 and nearing GPT-4's score of 9.26. Qualitative evaluations further demonstrate SparklesChat's generality in handling real-world applications. All resources are available at 

Generative AI has the potential to transform industries and society, but responsible design and collaboration among stakeholders are crucial.

The "Responsible AI Leadership: A Global Summit on Generative AI" was held in April 2023 to guide experts and policymakers in developing and governing generative AI systems responsibly. Over 100 thought leaders and practitioners participated, discussing key recommendations for responsible development, open innovation, and social progress. These 30 action-oriented recommendations aim to navigate AI complexities and harness its potential ethically. By implementing them, we can shape a more innovative, equitable, and prosperous future while mitigating risks.

Generative AI systems across modalities, ranging from text, image, audio, and video, have broad social impacts, but there exists no official standard for means of evaluating those impacts and which impacts should be evaluated. We move toward a standard approach in evaluating a generative AI system for any modality, in two overarching categories: what is able to be evaluated in a base system that has no predetermined application and what is able to be evaluated in society. We describe specific social impact categories and how to approach and conduct evaluations in the base technical system, then in people and society. Our framework for a base system defines seven categories of social impact: bias, stereotypes, and representational harms; cultural values and sensitive content; disparate performance; privacy and data protection; financial costs; environmental costs; and data and content moderation labor costs. Suggested methods for evaluation apply to all modalities and analyses of the limitations of existing evaluations serve as a starting point for necessary investment in future evaluations. We offer five overarching categories for what is able to be evaluated in society, each with their own subcategories: trustworthiness and autonomy; inequality, marginalization, and violence; concentration of authority; labor and creativity; and ecosystem and environment. Each subcategory includes recommendations for mitigating harm. We are concurrently crafting an evaluation repository for the AI research community to contribute existing evaluations along the given categories. This version will be updated following a CRAFT session at ACM FAccT 2023.

How do we benchmark the social impact of generative systems? This cross-institutional research collaboration provides a guide to evaluating the social impact of Generative AI Systems.

Emergent properties have been widely adopted as a term to describe behavior not present in smaller models but observed in larger models. Recent work suggests that the trade-off incurred by quantization is also an emergent property, with sharp drops in performance in models over 6B parameters. In this work, we ask "are quantization cliffs in performance solely a factor of scale?" Against a backdrop of increased research focus on why certain emergent properties surface at scale, this work provides a useful counter-example. We posit that it is possible to optimize for a quantization friendly training recipe that suppresses large activation magnitude outliers. Here, we find that outlier dimensions are not an inherent product of scale, but rather sensitive to the optimization conditions present during pre-training. This both opens up directions for more efficient quantization, and poses the question of whether other emergent properties are inherent or can be altered and conditioned by optimization and architecture design choices. We successfully quantize models ranging in size from 410M to 52B with minimal degradation in performance.

Large Language Models (LLMs) can achieve strong performance on many tasks by producing step-by-step reasoning before giving a final output, often referred to as chain-of-thought reasoning (CoT). It is tempting to interpret these CoT explanations as the LLM's process for solving a task. However, we find that CoT explanations can systematically misrepresent the true reason for a model's prediction. We demonstrate that CoT explanations can be heavily influenced by adding biasing features to model inputs -- e.g., by reordering the multiple-choice options in a few-shot prompt to make the answer always "(A)" -- which models systematically fail to mention in their explanations. When we bias models toward incorrect answers, they frequently generate CoT explanations supporting those answers. This causes accuracy to drop by as much as 36% on a suite of 13 tasks from BIG-Bench Hard, when testing with GPT-3.5 from OpenAI and Claude 1.0 from Anthropic. On a social-bias task, model explanations justify giving answers in line with stereotypes without mentioning the influence of these social biases. Our findings indicate that CoT explanations can be plausible yet misleading, which risks increasing our trust in LLMs without guaranteeing their safety. CoT is promising for explainability, but our results highlight the need for targeted efforts to evaluate and improve explanation faithfulness.

The BigScience Workshop was a value-driven initiative that spanned one and half years of interdisciplinary research and culminated in the creation of ROOTS, a 1.6TB multilingual dataset that was used to train BLOOM, one of the largest multilingual language models to date. In addition to the technical outcomes and artifacts, the workshop fostered multidisciplinary collaborations around large models, datasets, and their analysis. This in turn led to a wide range of research publications spanning topics from ethics to law, data governance, modeling choices and distributed training. This paper focuses on the collaborative research aspects of BigScience and takes a step back to look at the challenges of large-scale participatory research, with respect to participant diversity and the tasks required to successfully carry out such a project. Our main goal is to share the lessons we learned from this experience, what we could have done better and what we did well. We show how the impact of such a social approach to scientific research goes well beyond the technical artifacts that were the basis of its inception.

Successful applications of deep reinforcement learning (deep RL) combine algorithmic design and careful hyper-parameter selection. The former often comes from iterative improvements over existing algorithms, while the latter is either inherited from prior methods or tuned for the specific method being introduced. Although critical to a method’s performance, the effect of the various hyper-parameter choices are often overlooked in favour of algorithmic advances. In this paper, we perform an initial empirical investigation into a number of often-overlooked hyperparameters for value-based deep RL agents, demonstrating their varying levels of importance. We conduct this study on a varied set of classic control environments which helps highlight the effect each environment has on an algorithm’s hyper-parameter sensitivity.

Perception of toxicity evolves over time and often differs between geographies and cultural backgrounds. Similarly, black-box commercially available APIs for detecting toxicity, such as the Perspective API, are not static, but frequently retrained to address any unattended weaknesses and biases. We evaluate the implications of these changes on the reproducibility of findings that compare the relative merits of models and methods that aim to curb toxicity. Our findings suggest that research that relied on inherited automatic toxicity scores to compare models and techniques may have resulted in inaccurate findings. Rescoring all models from HELM, a widely respected living benchmark, for toxicity with the recent version of the API led to a different ranking of widely used foundation models. We suggest caution in applying apples-to-apples comparisons between studies and lay recommendations for a more structured approach to evaluating toxicity over time. Code and data are available at 

How changes to the Perspective API, used widely for toxicity evaluation, impact research reproducibility and rankings of model risk.

Hyperparameter optimization (HPO) and neural architecture search (NAS) are methods of choice to obtain the best-in-class machine learning models, but in practice they can be costly to run. When models are trained on large datasets, tuning them with HPO or NAS rapidly becomes prohibitively expensive for practitioners, even when efficient multi-fidelity methods are employed. We propose an approach to tackle the challenge of tuning machine learning models trained on large datasets with limited computational resources. Our approach, named PASHA, extends ASHA and is able to dynamically allocate maximum resources for the tuning procedure depending on the need. The experimental comparison shows that PASHA identifies well-performing hyperparameter configurations and architectures while consuming significantly fewer computational resources than ASHA.

Recent work in natural language processing (NLP) has yielded appealing results from scaling model parameters and training data; however, using only scale to improve performance means that resource consumption also grows. Such resources include data, time, storage, or energy, all of which are naturally limited and unevenly distributed. This motivates research into efficient methods that require fewer resources to achieve similar results. This survey synthesizes and relates current methods and findings in efficient NLP. We aim to provide both guidance for conducting NLP under limited resources, and point towards promising research directions for developing more efficient methods.

Synthesizing methods and findings in NLP efficiencies, guiding new researchers in the field, and inspiring the development of new methods.

Text embeddings are commonly evaluated on a small set of datasets from a single task not covering their possible applications to other tasks. It is unclear whether state-of-the-art embeddings on semantic textual similarity (STS) can be equally well applied to other tasks like clustering or reranking. This makes progress in the field difficult to track, as various models are constantly being proposed without proper evaluation. To solve this problem, we introduce the Massive Text Embedding Benchmark (MTEB). MTEB spans 8 embedding tasks covering a total of 58 datasets and 112 languages. Through the benchmarking of 33 models on MTEB, we establish the most comprehensive benchmark of text embeddings to date. We find that no particular text embedding method dominates across all tasks. This suggests that the field has yet to converge on a universal text embedding method and scale it up sufficiently to provide state-of-the-art results on all embedding tasks. MTEB comes with open-source code and a public leaderboard at 

We propose EventFormer– a computationally efficient event-based representation learning framework for asynchronously processing event camera data. EventFormer treats sparse input events as a spatially unordered set and models their spatial interactions using self-attention mechanism. An associative memory augmented recurrent module is used to correlate with the stored representation computed from past events. A memory addressing mechanism is proposed to store and retrieve the latent states only where these events occur and update them only when they occur. The representation learning shift from input space to the latent memory space resulting in reduced computation cost for processing each event. We show that EventFormer achieves 0.5% and 9% better accuracy with 30000× and 200× less computation compared to the state-of-the-art dense and event-based method, respectively, on event-based object recognition datasets.

Ensembling independent deep neural networks (DNNs) is a simple and effective way to improve top-line metrics and to outperform larger single models. In this work, we go beyond top-line metrics and instead explore the impact of ensembling on subgroup performances. Surprisingly, even with a simple homogenous ensemble – all the individual models share the same training set, architecture, and design choices – we find compelling and powerful gains in worst-k and minority group performance, i.e. fairness naturally emerges from ensembling. We show that the gains in performance from ensembling for the minority group continue for far longer than for the majority group as more models are added. Our work establishes that simple DNN ensembles can be a powerful tool for alleviating disparate impact from DNN classifiers, thus curbing algorithmic harm. We also explore why this is the case. We find that even in homogeneous ensembles, varying the sources of stochasticity through parameter initialization, mini-batch sampling, and the data-augmentation realizations, results in different fairness outcomes.

Multilingual models are often particularly dependent on scaling to generalize to a growing number of languages. Compression techniques are widely relied upon to reconcile the growth in model size with real world resource constraints, but compression can have a disparate effect on model performance for low-resource languages. It is thus crucial to understand the trade-offs between scale, multilingualism, and compression. In this work, we propose an experimental framework to characterize the impact of sparsifying multilingual pre-trained language models during fine-tuning. Applying this framework to mBERT named entity recognition models across 40 languages, we find that compression confers several intriguing and previously unknown generalization properties. In contrast to prior findings, we find that compression may improve model robustness over dense models. We additionally observe that under certain sparsification regimes compression may aid, rather than disproportionately impact the performance of low-resource languages.

Exploring compression as a way to improve model robustness for low-resource languages.

In the past few years, neural architecture search (NAS) has become an increasingly important tool within the deep learning community. Despite the many recent successes of NAS, however, most existing approaches operate within highly structured design spaces, and hence explore only a small fraction of the full search space of neural architectures while also requiring significant manual effort from domain experts. In this work, we develop techniques that enable efficient NAS in a significantly larger design space. To accomplish this, we propose to perform NAS in an abstract search space of program properties. Our key insights are as follows: (1) the abstract search space is significantly smaller than the original search space, and (2) architectures with similar program properties also have similar performance; thus, we can search more efficiently in the abstract search space. To enable this approach, we also propose a novel efficient synthesis procedure, which accepts a set of promising program properties, and returns a satisfying neural architecture. We implement our approach, αNAS, within an evolutionary framework, where the mutations are guided by the program properties. Starting with a ResNet-34 model, αNAS produces a model with slightly improved accuracy on CIFAR-10 but 96% fewer parameters. On ImageNet, αNAS is able to improve over Vision Transformer (30% fewer FLOPS and parameters), ResNet-50 (23% fewer FLOPS, 14% fewer parameters), and EfficientNet (7% fewer FLOPS and parameters) without any degradation in accuracy.

Despite widespread use of LLMs as conversational agents, evaluations of performance fail to capture a crucial aspect of communication: interpreting language in context. Humans interpret language using beliefs and prior knowledge about the world. For example, we intuitively understand the response "I wore gloves" to the question "Did you leave fingerprints?" as meaning "No". To investigate whether LLMs have the ability to make this type of inference, known as an implicature, we design a simple task and evaluate widely used state-of-the-art models. We find that, despite only evaluating on utterances that require a binary inference (yes or no), most perform close to random. Models adapted to be "aligned with human intent" perform much better, but still show a significant gap with human performance. We present our findings as the starting point for further research into evaluating how LLMs interpret language in context and to drive the development of more pragmatic and useful models of human discourse.

Investigating the implicature gap in Large Language Models.

Recent work has shown that augmenting environments with language descriptions improves policy learning. However, for environments with complex language abstractions, learning how to ground language to observations is difficult due to sparse, delayed rewards. We propose Language Dynamics Distillation (LDD), which pretrains a model to predict environment dynamics given demonstrations with language descriptions, and then fine-tunes these language-aware pretrained representations via reinforcement learning (RL). In this way, the model is trained to both maximize expected reward and retain knowledge about how language relates to environment dynamics. On SILG, a benchmark of five tasks with language descriptions that evaluate distinct generalization challenges on unseen environments (NetHack, ALFWorld, RTFM, Messenger, and Touchdown), LDD outperforms tabula-rasa RL, VAE pretraining, and methods that learn from unlabeled demonstrations in inverse RL and reward shaping with pretrained experts. In our analyses, we show that language descriptions in demonstrations improve sample-efficiency and generalization across environments, and that dynamics modelling with expert demonstrations is more effective than with non-experts.

Training deep neural networks in low rank, i.e. with factorised layers, is of particular interest to the community: it offers efficiency over unfactorised training in terms of both memory consumption and training time. Prior work has focused on low rank approximations of pre-trained networks and training in low rank space with additional objectives, offering various ad hoc explanations for chosen practice. We analyse techniques that work well in practice, and through extensive ablations on models such as GPT2 we provide evidence falsifying common beliefs in the field, hinting in the process at exciting research opportunities that still need answering.

Training on web-scale data can take months. But most computation and time is wasted on redundant and noisy points that are already learnt or not learnable. To accelerate training, we introduce Reducible Holdout Loss Selection (RHO-LOSS), a simple but principled technique which selects approximately those points for training that most reduce the model's generalization loss. As a result, RHO-LOSS mitigates the weaknesses of existing data selection methods: techniques from the optimization literature typically select 'hard' (e.g. high loss) points, but such points are often noisy (not learnable) or less task-relevant. Conversely, curriculum learning prioritizes 'easy' points, but such points need not be trained on once learned. In contrast, RHO-LOSS selects points that are learnable, worth learning, and not yet learnt. RHO-LOSS trains in far fewer steps than prior art, improves accuracy, and speeds up training on a wide range of datasets, hyperparameters, and architectures (MLPs, CNNs, and BERT). On the large web-scraped image dataset Clothing-1M, RHO-LOSS trains in 18x fewer steps and reaches 2% higher final accuracy than uniform data shuffling.

Modern machine learning research relies on relatively few carefully curated datasets. Even in these datasets, and typically in ‘untidy’ or raw data, practitioners are faced with significant issues of data quality and diversity which can be prohibitively labor intensive to address. Existing methods for dealing with these challenges tend to make strong assumptions about the particular issues at play, and often require a priori knowledge or metadata such as domain labels. Our work is orthogonal to these methods: we instead focus on providing a unified and efficient framework for Metadata Archaeology – uncovering and inferring metadata of examples in a dataset. We curate different subsets of data that might exist in a dataset (e.g. mislabeled, atypical, or out-of-distribution examples) using simple transformations, and leverage differences in learning dynamics between these probe suites to infer metadata of interest. Our method is on par with far more sophisticated mitigation methods across different tasks: identifying and correcting mislabeled examples, classifying minority-group samples, prioritizing points relevant for training and enabling scalable human auditing of relevant examples.

Providing a unified and efficient framework for Metadata Archaeology – uncovering and inferring metadata of examples in a dataset.

The number of parameters in state of the art neural networks has drastically increased in recent years. This surge of interest in large scale neural networks has motivated the development of new distributed training strategies enabling such models. One such strategy is model-parallel distributed training. Unfortunately, model-parallelism can suffer from poor resource utilisation, which leads to wasted resources. In this work, we improve upon recent developments in an idealised model-parallel optimisation setting: local learning. Motivated by poor resource utilisation in the global setting and poor task performance in the local setting, we introduce a class of intermediary strategies between local and global learning referred to as interlocking backpropagation. These strategies preserve many of the compute-efficiency advantages of local optimisation, while recovering much of the task performance achieved by global optimisation. We assess our strategies on both image classification ResNets and Transformer language models, finding that our strategy consistently out-performs local learning in terms of task performance, and out-performs global learning in training efficiency.

We study the impact of different pruning techniques on the representation learned by deep neural networks trained with contrastive loss functions. Our work finds that at high sparsity levels, contrastive learning results in a higher number of misclassified examples relative to models trained with traditional cross-entropy loss. To understand this pronounced difference, we use metrics such as the number of PIEs (Hooker et al., 2019), Q-Score (Kalibhat et al., 2022) and PDScore (Baldock et al., 2021) to measure the impact of pruning on the learned representation quality. Our analysis suggests the schedule of the pruning method implementation matters. We find that the negative impact of sparsity on the quality of the learned representation is the highest when pruning is introduced early-on in training phase.

Knowledge distillation has proven to be an effective technique in improving the performance a student model using predictions from a teacher model. However, recent work has shown that gains in average efficiency are not uniform across subgroups in the data, and in particular can often come at the cost of accuracy on rare subgroups and classes. To preserve strong performance across classes that may follow a long-tailed distribution, we develop distillation techniques that are tailored to improve the student’s worst-class performance. Specifically, we introduce robust optimization objectives in different combinations for the teacher and student, and further allow for training with any tradeoff between the overall accuracy and the robust worst-class objective. We show empirically that our robust distillation techniques not only achieve better worst-class performance, but also lead to Pareto improvement in the tradeoff between overall performance and worstclass performance compared to other baseline methods. Theoretically, we provide insights into what makes a good teacher when the goal is to train a robust student.

Modern large language models require distributed training strategies due to their size. The challenges of efficiently and robustly training them are met with rapid developments on both software and hardware frontiers. In this technical report, we explore challenges and design decisions associated with developing a scalable training framework, and present a quantitative analysis of efficiency improvements coming from adopting new software and hardware solutions.

Language models trained on large-scale unfiltered datasets curated from the open web acquire systemic biases, prejudices, and harmful views from their training data. We present a methodology for programmatically identifying and removing harmful text from web-scale datasets. A pretrained language model is used to assess the loglikelihood of researcher-written trigger phrases conditioned on a specific document, which is used to identify and filter documents from the dataset. We demonstrate that models trained on this filtered dataset exhibit lower propensity to generate harmful text, with a marginal decrease in performance on standard language modeling benchmarks compared to unfiltered baselines. We provide a partial explanation for this performance gap by surfacing examples of hate speech and other undesirable content from standard language modeling benchmarks. Finally, we discuss the generalization of this method and how trigger phrases reflecting specific values can be used by researchers to build language models which are more closely aligned with their values.

The One Billion Word Benchmark is a dataset derived from the WMT 2011 News Crawl, commonly used to measure language modeling ability in natural language processing. We train models solely on Common Crawl web scrapes partitioned by year, and demonstrate that they perform worse on this task over time due to distributional shift. Analysis of this corpus reveals that it contains several examples of harmful text, as well as outdated references to current events. We suggest that the temporal nature of news and its distribution shift over time makes it poorly suited for measuring language modeling ability, and discuss potential impact and considerations for researchers building language models and evaluation datasets.

Twitter has become one of the main information sharing platforms for millions of users world-wide. Numerous tweets are created daily, many with highly time sensitive content such as breaking news, new multimedia content or personal updates. Consequently, accurately recommending relevant tweets to users in a timely manner is a highly important and challenging problem. The 2020 ACM RecSys Challenge is aimed at benchmarking leading recommendation models for this task. The challenge is based on a large and recent dataset of over 200M tweet engagements released by Twitter with content in over 50 languages. In this work we present our approach where we leverage recent advances in deep language modeling and attention architectures, to combine information from extracted features, user engagement history and target tweet content. We first fine-tune leading multilingual language models M-BERT and XLM-R for Twitter data. Embeddings from these models are used to extract tweet and user history representations. We then combine all components together and jointly train them to maximize engagement prediction accuracy. Our approach achieves highly competitive performance placing 2’nd on the final private leaderboard. Full code is available here: https://github.com/layer6ai-labs/RecSys2020.

Research papers

Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models

The PRISM Alignment Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models

From One to Many: Expanding the Scope of Toxicity Mitigation in Language Models

Here's a Free Lunch: Sanitizing Backdoored Models with Model Merge

Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs

Elo Uncovered: Robustness and Best Practices in Language Model Evaluation

Which Prompts Make The Difference? Data Prioritization For Efficient Human LLM Evaluation

Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Models

Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning

When Less is More: Investigating Data Pruning for Pretraining LLMs at Scale

Sparkles: Unlocking Chats Across Multiple Images for Multimodal Instruction-Following Models

Intriguing Properties of Quantization at Scale

Efficient Methods for Natural Language Processing: A Survey

Intriguing Properties of Compression on Multilingual Models

Large Language Models are not Zero Shot Communicators

Predicting Twitter Engagement With Deep Language Models