18736 research outputs found
Sort by
Rethinking Retrieval Augmented Fine-Tuning in an evolving LLM landscape
Get PDFThis study explores the utilization of Retrieval Augmented Fine-Tuning (RAFT) to enhance the performance of Large Language Models (LLMs) in domain-specific Retrieval Augmented Generation (RAG) tasks. By integrating domain-specific information during the retrieval process, RAG aims to reduce hallucination and improve the accuracy of LLM outputs. We investigate the use of RAFT, an approach that enhances LLMs by incorporating domain-specific knowledge and effectively handling distractor documents. This paper validates previous work, which found that RAFT can considerably improve the performance of Llama2-7B in specific domains. We also expand upon previous work into new state-of-the-art open-source models and other datasets with mixed results. After fine-tuning three models (Llama2-7B, Llama3-8B, and Mistral-7B-v0.3) using RAFT and evaluating their performance compared to their instruction-tuned versions, our results suggest that RAFT can only improve accuracy for older LLMs on domain specific data. This effect was not found in the latest generation of open-source LLMs
Infringement Episodes
Get PDFFor decades, copyright scholars have waged a spirited campaign against statutory damages. Our remedial system, critics say, is an incoherent mess. The core problem is that copyright holders can recover a separate award of statutory damages for every infringed work. As a result, damages can rapidly add up in any case involving multiple works. Because the number of statutory awards is tethered to the number of works, even trivial claims can lead to crippling damages. Commentators, policymakers, and judges have criticized this system as arbitrary and overbroad. And yet it endures. This Article argues that copyright’s per-work scheme has obscured, and at times eclipsed, a more compelling paradigm of copyright damages—one that attends more closely to the defendant’s course of conduct. This new approach would allow courts to examine whether the defendant’s actions arose out of, and were rooted in, a single infringement episode. By infringement episode, I mean a chain of infringing acts that together constitute a larger factual event. When the defendant’s conduct is traceable to a single episode, courts should issue only a single statutory award—no matter how many works are at stake. This framework, in short, would substitute rigidity for flexibility. It would displace copyright’s one-award-per-work scheme and instead introduce a contextual inquiry into the defendant’s course of conduct. Doing so can mitigate the risk of outlandish awards, encourage courts to properly calibrate damages, and infuse a degree of much-needed pragmatism into our system
Investigating a Pollinator Curriculum (1-3 grade)
Get PDFIn this unit, elementary graders will identify and explain the pollen, flowering plants, and pollination process, and labelled various parts of a flower, pollinators. They will build the model of pollinators, initial models of the reproductive parts of plants. They will explore the 3D shapes, magnifying glasses, area and volume, modelling and explain low-high fidelity prototyping, engineering design process. The unit is designed by providing an abundance of practical ways to learn the significance of the pollination process and to get to know about pollinators and pollen in a place-based environment. Here, students will engage, explore, explain, elaborate and evaluate the new knowledge by applying hands-on engineering practices, technology practices, model-based learning, foster fostering communication, critical thinking, collaborating and creativity skills to solve the real-life problem which targeting their epistemic, conceptual, and social learning goals
Provenance Mining based Recommendation for Scientific Workflow Composition
No full textIn today\u27s web landscape, the vast number of available reusable and universally accessible web services, or so-called Application Programming Interfaces (APIs), facilitates the creation of big data analytics procedures (scientific workflows or workflows in short, or mashups). However, significant challenges are also posed due to the overwhelming variety of options of service candidates to choose from. A manual service selection process is often time-consuming and prone to errors, making it difficult to align service choices with user intentions efficiently. My research addresses this challenge by leveraging machine learning techniques to enhance the accuracy, speed, and robustness of service recommendations. My major contributions are three-fold. First, a novel knowledge graph framework called Unit of Work (UoW) knowledge graph (KG) is established, which records service dependencies within and among workflows. Second, service features are treated as first-class entities, transforming low-order UoW KGs into higher-order modal KGs. Third, based on the UoW networks, a generative adversarial network (GAN)-powered recommend-as-you-go approach is developed capable of predicting user intent and suggesting the next suitable services during a workflow construction process. Extensive experiments over real-world datasets, such as ProgrammableWeb.com and MyExperiment.org, have demonstrated the effectiveness and efficiency of our techniques. My techniques have also been integrated into the NASA EcoPro project
Microwell Fabrication for Impedance and AC Measurements of Isolated Ventricular Cardiomyocytes
Get PDFElectrophysiological behavior of human hearts have been extensively studied for hundreds of years. Even with elucidation of not just biological function, but also time and frequency-dependent current flow, the sheer intricacy of in vivo cardiac tissue function remains somewhat enigmatic. All while heart disease has been the leading cause of death in the United States for over a century, with its rates of occurrence still increasing annually [1]. Thus, new research techniques and tools to quantify cardiac function are still being pursued at smaller and smaller scales to decrease research/innovation time, cost, and diagnostic ambiguity. In fact, it remains incredibly scientifically-relevant to measure ionic current from single, isolated cells to quantify cells’ coupled electromechanical behavior. Among engineering studies, it is commonplace to use cultures of induced pluripotent STEM cells (iPSCs) to conveniently and uniformly mimic the behavior of a specific native cell type [2,3]. However, live cardiac cells from humans or animal models are non-uniform with exceptional capabilities of adaptation to varying electrical input (i.e., changing heart rate). This thesis outlines a novel microelectrode designed to measure impedance profiles for native murine left ventricular myocytes (LVMs) as utilized with cultured iPSC models.
The main goal of this thesis is to synthesize and test this microelectrode optimized for quantifying impedance of hundreds of both living and dead LVMs in real-time. The results can be employed in many ways including, but not limited to, quantification of cell-electrode adhesion, cell-cell contacts, cell death, contraction rates, cardiotoxicity of drugs, and cell morphology
Intelligent Solutions for Retroactive Anomaly Detection and Resolution with Log File Systems
Get PDFThis paper explores the intricate challenges log files pose from data science and machine learning perspectives. Drawing inspiration from existing methods, LAnoBERT, PULL, LLMs, and the breadth of recent research, this paper aims to push the boundaries of machine learning for log file systems. Our study comprehensively examines the unique challenges presented in our problem setup, delineates the limitations of existing methods, and introduces innovative solutions. These contributions are organized to offer valuable insights, predictions, and actionable recommendations tailored for Microsoft\u27s engineers working on log data analysis
Beyond the Horizon: Exploring Anomaly Detection Potentials with Federated Learning and Hybrid Transformers in Spacecraft Telemetry
Get PDFTelemetry sensors play a crucial role in spacecraft operations, providing essential data on efficiency, sustainability, and safety. However, identifying irregularities in telemetry data can be a time-consuming process that risks the success of missions. With the rise of CubeSats and smallsats, telemetry data has become more abundant, but concerns about privacy and scalability have resulted in untapped data potential. To address these issues, we propose a new approach to anomaly detection that utilizes machine learning models at data sources. These models solely transmit weights to a centralized server for aggregation, resulting in improved dataset performance with a single global model. We have also incorporated self-attention into the federated process to further enhance anomaly detection performance. Our experiments with real-world telemetry data have demonstrated that our approach is state-of-the-art in that we can construct a single model to address multiple telemetry channels while still adhering to the constraints typically seen in space missions. Our framework streamlines anomaly detection, promoting operational efficiency, sustainability, and safety. It facilitates collaborative insights while abiding by mission security constraints and reducing the risk of accidents and downtime, ensuring sustainability
Reconciling Domestic Violence Protections and the Second Amendment
Get PDFIn March of 2023, the Fifth Circuit Court of Appeals held that individuals subject to domestic violence protective orders could not be required to give up their guns. The decision was the first of a federal court to overturn a firearm regulation pursuant to New York State Rifle & Pistol Association v. Bruen, a 2022 Supreme Court opinion that created a new standard for determining the constitutionality of gun restrictions. After Bruen, only laws that are “consistent with this Nation’s historical tradition of firearm regulation” pass constitutional muster.The Fifth’s Circuit decision in U.S. v. Rahimi, which the Supreme Court will review in the 2023-24 term, highlights the unworkability of the Bruen test. Women’s rights were virtually nonexistent when the Second Amendment was ratified. Domestic violence was tolerated, and it was not until nearly 200 years later that protective order statutes were enacted across the United States. Looking to the past to justify modern-day gun safety laws gravely threatens women’s rights and safety.But Bruen does not require such a narrow reading. Significant historical and legal precedent exists for disarming dangerous persons, and those who have had protective orders entered against them undoubtedly fall into that category. This article’s feminist critique of Bruen demonstrates why its holding is deeply problematic, but it also shows that it is possible to both hew to Second Amendment jurisprudence and protect survivors of intimate partner violence
Experiments on Employer-Employee Relationships
Get PDFEmployer-employee relationships exist commonly in workplaces and are important for people to understand the interactions between employers and employees and also their own behavior. This dissertation presents studies of different aspects of employer-employee relationships using laboratory experiments. Chapter 1 studies how the problematic nature of many middle management positions drive middle managers to interact with their employees in a negative way in workplaces, as they are under pressure from upper management to extract high effort from their employees, but given few incentives to do so. Chapter 2 focuses on the potential peer effects of employees\u27 performance feedback and performance. My current data suggests that there is no overall treatment effect of peer workers\u27 performance feedback or performance on employees\u27 effort choices, but a positive correlation between coworkers\u27 efforts and employees\u27 efforts is found and indicates the peer effects of performance can potentially exist, which echos the results from a number of studies about peer effects. Chapter 3 examines if females and minorities have reduced willingness to participate in tasks with more subjective judgment. My data provides suggestive but not significant results to support my hypotheses that correlations between subjects\u27 avoidance to subjective judgment and their genders, ethnicities, or demographic information being accessible to the judge exist. Correlations between subjects\u27 changes of avoidance to judgment and their genders or ethnicities are found insignificant as well