36214 research outputs found
Sort by
Magpie: Generating High-Quality Synthetic Data with Open-Source Large Language Models
No full textThesis (Master's)--University of Washington, 2025High-quality instruction data is critical for aligning large language models (LLMs). Although some models, such as Llama-3-Instruct, have open weights, their alignment data remain private, which hinders the democratization of AI. High human labor costs and a limited, predefined scope for prompting prevent existing open-source data creation methods from scaling effectively, potentially limiting the diversity and quality of public alignment datasets.Is it possible to synthesize high-quality instruction data at scale by extracting it directly from an aligned LLM? We present a self-synthesis method for generating large-scale alignment data named Magpie. Our key observation is that aligned LLMs like Llama-3-Instruct can generate a user query when we input only the pre-query templates up to the position reserved for user messages, thanks to their auto-regressive nature. We use this method to prompt Llama-3-Instruct and generate 4 million instructions along with their corresponding responses. We further introduce extensions of Magpie for filtering, generating multi-turn, preference optimization, domain-specific and multilingual datasets. We perform a comprehensive analysis of the Magpie-generated data. To compare Magpie-generated data with other public instruction datasets (e.g., ShareGPT, WildChat, Evol-Instruct, UltraChat, OpenHermes, Tulu-V2-Mix, GenQA), we fine-tune Llama-3-8B-Base with each dataset and evaluate the performance of the fine-tuned models. Our results indicate that using Magpie for supervised fine-tuning (SFT) solely can surpass the performance of previous public datasets utilized for both SFT and preference optimization, such as direct preference optimization with UltraFeedback. We also show that in some tasks, models supervised fine-tuned with Magpie perform comparably to the official Llama-3-8B-Instruct, despite the latter being enhanced with 10 million data points through SFT and subsequent preference optimization. This advantage is evident on alignment benchmarks such as AlpacaEval, ArenaHard, and WildBench
Chutzpah: The Jewish Voices Behind American Comics
No full textThesis (Master's)--University of Washington, 2025A companion to the Chutzpah: The Jewish Voices Behind American Comics exhibit at the Museum of Pop Culture (MoPOP), this curator's guide elaborates on a selection of key Jewish Marvel characters as seen at MoPOP while expanding on the exhibit's themes of Jewish identity and resilience. Since their creation, comic books have been a powerful tool in combating hatred and they continue to be today along with creating a safe space for Jewish readers. The guide contains an essay on American antisemitism during Marvel's golden and silver eras (1930-1960s), how Marvel's Jewish staff were responding to it, and how that same antisemitism has carried through to the present. In looking at contemporary antisemitism the power of comic books to combat antisemitism and promote empathy is examined
Exploring the Impact of Patient Navigation on Colorectal Cancer Screening Rates among Clinics in Washington State
No full textThesis (Master's)--University of Washington, 2025Introduction: The Centers for Disease Control and Prevention (CDC) created the Colorectal Cancer Control Program (CRCCP) to increase colorectal cancer (CRC) screening uptake. The CRCCP partners with primary care clinics to increase the implementation of evidence-based interventions (EBIs) and supporting activities (SAs) such as patient navigation (PN) in clinics. This thesis assesses the impact of newly implementing PN compared to continuing or enhancing PN on the change in CRC screening rate.Methods: Data from clinics participating in Washington state’s CRCCP from July 2022 to June 2024 was used to determine clinic characteristics and change in screening rate from program year (PY) 3 to PY4. Clinics were stratified into two groups: those that had newly implemented PN vs those that had continued or enhanced PN. I calculated descriptive statistics for clinic characteristics such as clinic size and percentage of uninsured patients stratified for each PN groups. Then I calculated the mean, median, range, and interquartile range for the change in CRC screening rate for each group and used a Mann-Whitney u-test to determine whether the distribution of the change in CRC screening rate was significantly different between the two groups.
Results: The PN new and PN continuing/enhancing groups differed greatly in clinic characteristics such as clinic size (72.73% large clinics vs. 20.00% large clinics, respectively) and percentage of uninsured patients (45.45% with ≤5% uninsured vs. 100.00% with ≤5% uninsured). The PN new in PY4 group had a much lower median change in CRC screening rate (-0.17%) compared to the PN continuing in PY4 group (14.30%). The difference in the distribution between these groups was statistically significant (p-value = 0.00795).
Discussion: Clinics that continued to implement or enhanced PN for more than one year experienced higher changes in CRC screening rate compared to those clinics that newly implemented PN.
Conclusion: PN is a promising intervention that can increase CRC screening rates over time. My analysis found that continuing or enhancing PN can lead to a greater increase in screening rates than newly implementing PN among CRCCP participating clinics in Washington state. Future research should focus on understanding how these clinics are enhancing PN and how possible confounders impact the relationship between PN and CRC screening rates
Waste is a Construct
No full textThesis (Master's)--University of Washington, 2025Waste is a Construct is an exhibit hosted at the Center for Wooden Boats located in Seattle, Washington, opening on June 20th, 2025. The pieces in the exhibit are made from derelict vessels, and are a combination of art and functional pieces. The main goals are to educate and bring light to the issue of derelict vessels and marine debris as a wider concept. Using art to explore the environmental and social impact that waste has on our planet and its inhabitants, it will begin to shed much needed light on the issues surrounding marine debris, reuse, and re-creation
Creative Aging at the Tacoma Art Museum.
No full textThesis (Master's)--University of Washington, 2025My thesis is a guideline of what is needed to plan and implement creative aging programs at the Tacoma art museum. It was created for the use by staff and volunteers involved in creative aging programs. It was developed using community input, pre existing guidelines for creative aging programs, and research on the outcomes of creative aging programs
Rent and the Dynamics of Housing Tenure, Search, and Policy: A Local Perspective from Brazilian Cities
No full textThesis (Ph.D.)--University of Washington, 2025Brazil’s persistent housing deficit, affecting approximately 6.2 million households in 2022, reflects longstanding limitations in the country's housing policy systems. Despite growing levels of cost burden, particularly among low-income renters, the provision of rental housing remains fragmented and underdeveloped, typically addressed through emergency programs or localized initiatives. As rental markets expand, both formally and informally, and as digital platforms transform the way people search for housing, understanding these shifts is crucial for informing equitable and sustainable housing policy. This dissertation investigates the changing landscape of rental housing in Brazil through three interrelated studies. The first examines the adoption and discontinuation of local-level policies between 2007 and 2019. It highlights how factors such as administrative capacity, mayoral turnover, and political ideology correlate with the stability of housing initiatives. The second study examines the geography and coverage of online rental platforms in the city of Belo Horizonte, uncovering spatial and socioeconomic disparities in listings that reflect and potentially perpetuate patterns of urban inequality. The third study examines the emergence of social rental housing policies in Brazilian municipalities, highlighting how these efforts are influenced by policy ambiguity, local resistance, and uneven implementation, despite growing interest in rent-based solutions. Together, these studies enhance our understanding of the growing importance of renting within Brazil's urban housing system, while offering a cautionary account of the volatility that characterizes housing policy in the country. It also critically assesses the promise and limitations of digital rental platforms as sources of data for housing research. In conclusion, this dissertation underscores the necessity of addressing the growing and shifting housing deficit through multifaceted solutions tailored to the diverse needs of low-income populations. It also offers insights intended to inform policy decisions based on evidence and the active, consistent involvement of affected communities
Prospective Memories: Futures of Monument Design
No full textThesis (Master's)--University of Washington, 2025As societies negotiate meaningfulness, their collective values are encapsulated by tangible artifacts produced through the activity of designing. Prospective Memories: Futures of Monument Design explores the role of design in shaping collective memory along with the resulting discourse generated by the monuments that inhabit our public spaces. These forms serve as physical embodiments of perceived significance and are designed with the intention to influence the value systems of future generations. Conventional design paradigms for monuments have emphasized permanence and unchangeability, leading to an inherited public memorial landscape that struggles to accommodate evolving community identities. This thesis proposes the re-evaluation of conventional monument design frameworks by using diegetic prototypes to explore prospective scenarios that transform monuments from static objects into active participants in cultural and political dialogues. Ultimately, this research aims to offer a point of reflection through which designers can critique the status quo, priming a future monument landscape that is better able to reflect and adapt to contemporary societal shifts
Stochastic Optimization in Disaster Operations Management: Medical Capacity Planning During Epidemic, Optimal Subsidy Policy For Renewal Technology Adoption and Data-driven Robust Supply Chain Optimization
No full textThesis (Ph.D.)--University of Washington, 2025Each of the problems addressed in this thesis falls under the broader umbrella of disaster operations management, where the focus is on minimizing the impact of unforeseen disruptions on critical systems. Whether it’s managing the surge in healthcare demand during a pandemic, promoting sustainable energy adoption to mitigate long-term environmental crises, or optimizing supply chains to withstand production disruptions, the core challenge remains the same: to design resilient systems capable of navigating uncertainty and resource constraints. These problems highlight the need for well-informed, strategic decision-making to ensure that essential services continue to function even in the face of crises. The first problem deals with managing medical equipment capacity during the early spread of an infection in a region, a critical research problem during COVID pandemic. After a brief introduction to infectious disease modeling, we develop a model for a regional decision- maker to analyze the requirement of medical equipment capacity in the early stages of a spread of infections. We use the model to propose and evaluate ways to manage limited equipment capacity. Early stage infection growth is captured by a stochastic differential equation (SDE) and is part of a two-period community spread and shutdown model. We use the running-maximum process of a geometric Brownian motion to develop a performance metric, probability of breach, for a given capacity level. Decision-maker estimates costs of economy versus health and the time till the availability of a cure; we develop a heuristic rule and an optimal formulation that use these estimates to determine the required medical equipment capacity. We connect the level of capacity to a menu of actions, including the level and timing of shutdown, shutdown effectiveness, and enforcement. Our results show how these actions can compensate for the limited medical equipment capacity in a region. We next address the sharing of medical equipment capacity across regions and its impact on the breach probability. In addition to traditional risk-pooling, we identify a peak-timing effect depending on when infections peak in different regions. We show that equipment sharing may not benefit the regions when capacity is tight. A coupled SDE model captures the messaging coordination and movement across regional borders. Numerical experiments on this model show that under certain conditions, such movement and coordination can synchronize the infection trajectories and bring the peaks closer, reducing the benefit of sharing capacity. We then turn our attention to the problem of devising optimal subsidy to encourage adoption of solar technology in a region. Each household in a population characterized by income heterogeneity faces random demand for electricity and decides if and when it should adopt a solar product, rooftop solar or community solar. A central planner, aiming to meet an adoption level target within a set time, offers net metering and subsidy on solar products and minimizes its total cost. Our focus is on analyzing the interactions of three new features we add to the literature: income diversity, availability of community solar, and consideration of adoption timing. We develop a bilevel optimization formulation to derive the optimal subsidy policy. The upper level (planner’s) problem is a constrained non-linear optimization model in which the planner aims to minimize the average subsidy cost. The lower level (household’s) problem is an optimal stopping formulation, which captures the adoption decisions of the households. We derive a closed-form expression for the distribution of optimal adoption time of households for a given subsidy policy. We show that the planner’s problem is convex in the case of homogeneous subsidy for the two products. Our results underscore the importance for planners to consider three factors - adoption level target, time target, and subsidy budget - simultaneously as they work in tandem to influence the adoption outcome. The planners must also consider the inclusion of community solar in their plans because, as we show, community and rooftop solar attract households from different sides of the income spectrum. In the presence of income inequality, the availability of community makes it easier to meet solar adoption targets.The third problem we consider is a classic mixed integer programming formulation for sup- ply chain optimization of a semi-conductor supplier in presence of supply disruptions. We provided a deterministic formulation for optimal outbound network routing. We integrate the past publicly available data in the optimization model and show the shortcomings of its optimal solution using a numerical experiment. We then designed a practical robust formulation to handle uncertain production capacity. We numerically study the optimal solutions for low and high tolerance to deviation from deterministic optimal solution and generate optimal production and transportation plan which is not only more resilient to random shocks but also requires marginal change from current plan
Biodiversity within Hedophyllum sessile
No full textAn observational study on whats in the holdfasts of Hedophyllum sessile.
Collecting four replicas of sea cabbages within a 1.5-foot ecology frame from two
locations, Dead Man’s Bay Preserve and Cattle Point. With these collected specimens
then looked over for any organisms found inside/on them and identified those
organisms found. The results of this observational study were that the highest amount
of a species was Lacuna vincta, being 56.4%, and an unknown species of egg colony
being 12.8%
Highly Multiplexed Fluorescence Microscopy with Spectrally Tunable Semiconducting Polymer Dots
No full textThesis (Ph.D.)--University of Washington, 2025The need for advanced imaging tools has grown rapidly with the increasing complexity of biological questions. To understand how cells behave and interact in near physiological conditions, in either thin tissue slices or across whole tissues, researchers must be able to map a wide range of molecular targets across space and scale. Traditional fluorescence imaging is limited to around five targets for a single sample. Commonly used probes often suffer from spectral overlap, narrow Stokes shifts, and susceptibility to photobleaching. This bottleneck restricts multiplexing and narrows the biological insight that can be gained from a single experiment. Beyond that, deep tissue imaging poses additional challenges. Light scattering and absorption, especially in thick or uncleared samples, can significantly degrade resolution and signal quality. Together, these limitations reduce both the depth and dimensionality of the information that can be extracted from complex specimens. As a result, questions involving spatial organization, tissue-scale signaling, or multicellular interactions often remain only partially answered. In this dissertation, I explore new strategies to push the boundaries of multiplexed fluorescence imaging, both in thin sections and in intact, cleared tissues. Chapter 1 introduces the core principles of fluorescence microscopy and highlights the constraints posed by conventional labeling tools in high-plex settings. Chapter 2 presents a new approach based on spectrally tunable semiconducting polymer dots (Pdots), which achieve large Stokes shifts and high brightness, enabling simultaneous detection of over 20 targets without specialized hardware or iterative cycles. Building on this foundation, Chapter 3 extends this capability into three dimensions through the integration of polymer-dot labeling with light sheet fluorescence microscopy and tissue clearing techniques. With further optimization on this system, I aim to build a fast, highly-multiplexed and scalable volumetric imaging platform, suitable for generating rich molecular and structural maps of biological systems of thick or whole organs