Human-produced emissions are growing at an alarming rate, causing already observable changes in the climate and environment in general. Each year global carbon dioxide emissions hit a new record, and it is reported that 0.5% of total US greenhouse gas emissions are attributed to data centres as of 2021. The release of ChatGPT in late 2022 sparked social interest in Large Language Models (LLMs), the new generation of Language Models with a large number of parameters and trained on massive amounts of data. Currently, numerous companies are releasing products featuring various LLMs, with many more models in development and awaiting release. Deep Learning research is a competitive field, with only models that reach top performance attracting attention and being utilized. Hence, achieving better accuracy and results is often the first priority, while the model's efficiency and the environmental impact of the study are neglected. However, LLMs demand substantial computational resources and are very costly to train, both financially and environmentally. It becomes essential to raise awareness and promote conscious decisions about algorithmic and hardware choices. Providing information on training time, the approximate carbon dioxide emissions and power consumption would assist future studies in making necessary adjustments and determining the compatibility of available computational resources with model requirements. In this study, we infused T5 LLM with external knowledge and fine-tuned the model for Question-Answering task. Furthermore, we calculated and reported the approximate environmental impact for both steps. The findings demonstrate that the smaller models may not always be sustainable options, and increased training does not always imply better performance. The most optimal outcome is achieved by carefully considering both performance and efficiency factors.

Generative AI holds the promise of enabling a range of sought-after capabilities and revolutionizing workflows in various consumer and enterprise verticals. However, putting a model in production involves much more than just generating an output. It involves ensuring the model is reliable, safe, performant and also adheres to the policy of operation in a particular domain. Guardrails as a necessity for models has evolved around the need to enforce appropriate behavior of models, especially when they are in production. In this paper, we use education as a use case, given its stringent requirements of the appropriateness of content in the domain, to demonstrate how a guardrail model can be trained and deployed in production. Specifically, we describe our experience in building a production-grade guardrail model for a K-12 educational platform. We begin by formulating the requirements for deployment to this sensitive domain. We then describe the training and benchmarking of our domain-specific guardrail model, which outperforms competing open- and closed- instruction-tuned models of similar and larger size, on proprietary education-related benchmarks and public benchmarks related to general aspects of safety. Finally, we detail the choices we made on architecture and the optimizations for deploying this service in production; these range across the stack from the hardware infrastructure to the serving layer to language model inference optimizations. We hope this paper will be instructive to other practitioners looking to create production-grade domain-specific services based on generative AI and large language models.

Although not all bots are malicious, the vast majority of them are responsible for spreading misinformation and manipulating the public opinion about several issues, i.e., elections and many more. Therefore, the early detection of bots is crucial. Although there have been proposed methods for detecting bots in social media, there are still substantial limitations. For instance, existing research initiatives still extract a large number of features and train traditional machine learning algorithms or use GloVe embeddings and train LSTMs. However, feature extraction is a tedious procedure demanding domain expertise. Also, language models based on transformers have been proved to be better than LSTMs. Other approaches create large graphs and train graph neural networks requiring in this way many hours for training and access to computational resources. To tackle these limitations, this is the first study employing only the user description field and images of three channels denoting the type and content of tweets posted by the users. Firstly, we create digital DNA sequences, transform them to 3d images, and apply pretrained models of the vision domain, including EfficientNet, AlexNet, VGG16, etc. Next, we propose a multimodal approach, where we use TwHIN-BERT for getting the textual representation of the user description field and employ VGG16 for acquiring the visual representation for the image modality. We propose three different fusion methods, namely concatenation, gated multimodal unit, and crossmodal attention, for fusing the different modalities and compare their performances. Finally, we present a qualitative analysis of the behavior of our best performing model. Extensive experiments conducted on the Cresci'17 and TwiBot-20 datasets demonstrate valuable advantages of our introduced approaches over state-of-the-art ones.

Large Language Models (LLMs) excel in diverse tasks but often underperform in specialized fields due to limited domain-specific or proprietary corpus. Continual pre-training (CPT) enhances LLM capabilities by imbuing new domain-specific or proprietary knowledge while replaying general corpus to prevent catastrophic forgetting. The data mixture ratio of general corpus and domain-specific corpus, however, has been chosen heuristically, leading to sub-optimal training efficiency in practice. In this context, we attempt to re-visit the scaling behavior of LLMs under the hood of CPT, and discover a power-law relationship between loss, mixture ratio, and training tokens scale. We formalize the trade-off between general and domain-specific capabilities, leading to a well-defined Critical Mixture Ratio (CMR) of general and domain data. By striking the balance, CMR maintains the model's general ability and achieves the desired domain transfer, ensuring the highest utilization of available resources. Therefore, if we value the balance between efficiency and effectiveness, CMR can be consider as the optimal mixture ratio.Through extensive experiments, we ascertain the predictability of CMR, and propose CMR scaling law and have substantiated its generalization. These findings offer practical guidelines for optimizing LLM training in specialized domains, ensuring both general and domain-specific performance while efficiently managing training resources.

Large Language Models (LLMs) have demonstrated impressive capabilities in generating diverse and contextually rich text. However, concerns regarding copyright infringement arise as LLMs may inadvertently produce copyrighted material. In this paper, we first investigate the effectiveness of watermarking LLMs as a deterrent against the generation of copyrighted texts. Through theoretical analysis and empirical evaluation, we demonstrate that incorporating watermarks into LLMs significantly reduces the likelihood of generating copyrighted content, thereby addressing a critical concern in the deployment of LLMs. Additionally, we explore the impact of watermarking on Membership Inference Attacks (MIAs), which aim to discern whether a sample was part of the pretraining dataset and may be used to detect copyright violations. Surprisingly, we find that watermarking adversely affects the success rate of MIAs, complicating the task of detecting copyrighted text in the pretraining dataset. Finally, we propose an adaptive technique to improve the success rate of a recent MIA under watermarking. Our findings underscore the importance of developing adaptive methods to study critical problems in LLMs with potential legal implications.

Machine unlearning, an emerging research topic focusing on compliance with data privacy regulations, enables trained models to remove the information learned from specific data. While many existing methods indirectly address this issue by intentionally injecting incorrect supervisions, they can drastically and unpredictably alter the decision boundaries and feature spaces, leading to training instability and undesired side effects. To fundamentally approach this task, we first analyze the changes in latent feature spaces between original and retrained models, and observe that the feature representations of samples not involved in training are closely aligned with the feature manifolds of previously seen samples in training. Based on these findings, we introduce a novel evaluation metric for machine unlearning, coined dimensional alignment, which measures the alignment between the eigenspaces of the forget and retain set samples. We employ this metric as a regularizer loss to build a robust and stable unlearning framework, which is further enhanced by integrating a self-distillation loss and an alternating training scheme. Our framework effectively eliminates information from the forget set and preserves knowledge from the retain set. Lastly, we identify critical flaws in established evaluation metrics for machine unlearning, and introduce new evaluation tools that more accurately reflect the fundamental goals of machine unlearning.

This paper introduces CovScore, an automatic reference-less methodology for evaluating thematic title sets, extracted from a corpus of documents. While such extraction methods are widely used, evaluating their effectiveness remains an open question. Moreover, some existing practices heavily rely on slow and laborious human annotation procedures. Inspired by recently introduced LLM-based judge methods, we propose a novel methodology that decomposes quality into five main metrics along different aspects of evaluation. This framing simplifies and expedites the manual evaluation process and enables automatic and independent LLM-based evaluation. As a test case, we apply our approach to a corpus of Holocaust survivor testimonies, motivated both by its relevance to title set extraction and by the moral significance of this pursuit. We validate the methodology by experimenting with naturalistic and synthetic title set generation systems and compare their performance with the methodology.

General-purpose artificial intelligence (AI) systems are built on massive swathes of public web data, assembled into corpora such as C4, RefinedWeb, and Dolma. To our knowledge, we conduct the first, large-scale, longitudinal audit of the consent protocols for the web domains underlying AI training corpora. Our audit of 14,000 web domains provides an expansive view of crawlable web data and how codified data use preferences are changing over time. We observe a proliferation of AI-specific clauses to limit use, acute differences in restrictions on AI developers, as well as general inconsistencies between websites' expressed intentions in their Terms of Service and their robots.txt. We diagnose these as symptoms of ineffective web protocols, not designed to cope with the widespread re-purposing of the internet for AI. Our longitudinal analyses show that in a single year (2023-2024) there has been a rapid crescendo of data restrictions from web sources, rendering ~5%+ of all tokens in C4, or 28%+ of the most actively maintained, critical sources in C4, fully restricted from use. For Terms of Service crawling restrictions, a full 45% of C4 is now restricted. If respected or enforced, these restrictions are rapidly biasing the diversity, freshness, and scaling laws for general-purpose AI systems. We hope to illustrate the emerging crises in data consent, for both developers and creators. The foreclosure of much of the open web will impact not only commercial AI, but also non-commercial AI and academic research.

As learning-to-rank models are increasingly deployed for decision-making in areas with profound life implications, the FairML community has been developing fair learning-to-rank (LTR) models. These models rely on the availability of sensitive demographic features such as race or sex. However, in practice, regulatory obstacles and privacy concerns protect this data from collection and use. As a result, practitioners may either need to promote fairness despite the absence of these features or turn to demographic inference tools to attempt to infer them. Given that these tools are fallible, this paper aims to further understand how errors in demographic inference impact the fairness performance of popular fair LTR strategies. In which cases would it be better to keep such demographic attributes hidden from models versus infer them? We examine a spectrum of fair LTR strategies ranging from fair LTR with and without demographic features hidden versus inferred to fairness-unaware LTR followed by fair re-ranking. We conduct a controlled empirical investigation modeling different levels of inference errors by systematically perturbing the inferred sensitive attribute. We also perform three case studies with real-world datasets and popular open-source inference methods. Our findings reveal that as inference noise grows, LTR-based methods that incorporate fairness considerations into the learning process may increase bias. In contrast, fair re-ranking strategies are more robust to inference errors. All source code, data, and experimental artifacts of our experimental study are available here: https://github.com/sewen007/hoiltr.git

In this essay, I argue that explicit ethical machines, whose moral principles are inferred through a bottom-up approach, are unable to replicate human-like moral reasoning and cannot be considered moral agents. By utilizing Alan Turing's theory of computation, I demonstrate that moral reasoning is computationally intractable by these machines due to the halting problem. I address the frontiers of machine ethics by formalizing moral problems into 'algorithmic moral questions' and by exploring moral psychology's dual-process model. While the nature of Turing Machines theoretically allows artificial agents to engage in recursive moral reasoning, critical limitations are introduced by the halting problem, which states that it is impossible to predict with certainty whether a computational process will halt. A thought experiment involving a military drone illustrates this issue, showing that an artificial agent might fail to decide between actions due to the halting problem, which limits the agent's ability to make decisions in all instances, undermining its moral agency.