Benchmark Data Research Articles

Working with benchmark datasets in the Cuby framework.

The development and benchmarking of computational chemistry methods rely on comparison with benchmark data. More and larger benchmark datasets are becoming available, and working efficiently with them is a necessity. The Cuby framework provides rich functionality for working with datasets, comes with many ready-to-use predefined benchmark sets, and interfaces with a wide range of computational chemistry software packages. Here, we review the tools Cuby provides for working with datasets and provide examples of more advanced workflows, such as handling large numbers of computations on high performance computing resources and reusing previously computed data. Cuby has also been extended recently to include two important benchmark databases, NCIAtlas and GMTKN55.

Developing a novel causal inference algorithm for personalized biomedical causal graph learning using meta machine learning

BackgroundModeling causality through graphs, referred to as causal graph learning, offers an appropriate description of the dynamics of causality. The majority of current machine learning models in clinical decision support systems only predict associations between variables, whereas causal graph learning models causality dynamics through graphs. However, building personalized causal graphs for each individual is challenging due to the limited amount of data available for each patient.MethodIn this study, we present a new algorithmic framework using meta-learning for learning personalized causal graphs in biomedicine. Our framework extracts common patterns from multiple patient graphs and applies this information to develop individualized graphs. In multi-task causal graph learning, the proposed optimized initial guess of shared commonality enables the rapid adoption of knowledge to new tasks for efficient causal graph learning.ResultsExperiments on one real-world biomedical causal graph learning benchmark data and four synthetic benchmarks show that our algorithm outperformed the baseline methods. Our algorithm can better understand the underlying patterns in the data, leading to more accurate predictions of the causal graph. Specifically, we reduce the structural hamming distance by 50-75%, indicating an improvement in graph prediction accuracy. Additionally, the false discovery rate is decreased by 20-30%, demonstrating that our algorithm made fewer incorrect predictions compared to the baseline algorithms.ConclusionTo the best of our knowledge, this is the first study to demonstrate the effectiveness of meta-learning in personalized causal graph learning and cause inference modeling for biomedicine. In addition, the proposed algorithm can also be generalized to transnational research areas where integrated analysis is necessary for various distributions of datasets, including different clinical institutions.

Gene expression programming (GEP) as novel tool for thermal analysis and kinetic modeling of pyrolysis reactions: coal pyrolysis case study

PurposeIn this study, a different approach is introduced to generate the kinetic sub-model for the modeling of solid-state pyrolysis reactions based on the thermogravimetric (TG) experimental data over a specified range of heating rates. Gene Expression Programming (GEP) is used to produce a correlation for the single-step global reaction rate as a function of determining kinetic variables, namely conversion, temperature, and heating rate.Design/methodology/approachFor a case study on the coal pyrolysis, a coefficient of determination (R2) of 0.99 was obtained using the generated model according to the experimental benchmark data. Comparison of the model results with the experimental data proves the applicability, reliability, and convenience of GEP as a powerful tool for modeling purposes in the solid-state pyrolysis reactions.FindingsThe resulting kinetic sub-model takes advantage of particular characteristics, to be highly efficient, simple, accurate, and computationally attractive, which facilitates the CFD simulation of real pyrolizers under isothermal and non-isothermal conditions.Originality/valueIt should be emphasized that the above-mentioned manuscript is not under evaluation in any journals and submitted exclusively for consideration for possible publication in this journal. The generated kinetic model is in the final form of an algebraic correlation which, in comparison to the conventional kinetic models, suggests several advantages: to be relatively simpler, more accurate, and numerically efficient. These characteristics make the proposed model computationally attractive when used as a sub-model in CFD applications to simulate real pyrolizers under complex heating conditions.

Classification of Major Solar Flares from Extremely Imbalanced Multivariate Time Series Data Using Minimally Random Convolutional Kernel Transform

Solar flares are characterized by sudden bursts of electromagnetic radiation from the Sun’s surface, and are caused by the changes in magnetic field states in active solar regions. Earth and its surrounding space environment can suffer from various negative impacts caused by solar flares, ranging from electronic communication disruption to radiation exposure-based health risks to astronauts. In this paper, we address the solar flare prediction problem from magnetic field parameter-based multivariate time series (MVTS) data using multiple state-of-the-art machine learning classifiers that include MINImally RandOm Convolutional KErnel Transform (MiniRocket), Support Vector Machine (SVM), Canonical Interval Forest (CIF), Multiple Representations Sequence Learner (Mr-SEQL), and a Long Short-Term Memory (LSTM)-based deep learning model. Our experiment is conducted on the Space Weather Analytics for Solar Flares (SWAN-SF) benchmark data set, which is a partitioned collection of MVTS data of active region magnetic field parameters spanning over nine years of operation of the Solar Dynamics Observatory (SDO). The MVTS instances of the SWAN-SF dataset are labeled by GOES X-ray flux-based flare class labels, and attributed to extreme class imbalance because of the rarity of the major flaring events (e.g., X and M). As a performance validation metric in this class-imbalanced dataset, we used the True Skill Statistic (TSS) score. Finally, we demonstrate the advantages of the MVTS learning algorithm MiniRocket, which outperformed the aforementioned classifiers without the need for essential data preprocessing steps such as normalization, statistical summarization, and class imbalance handling heuristics.

Constrained Device Performance Benchmarking with the Implementation of Post-Quantum Cryptography

Advances in quantum computers may pose a significant threat to existing public-key encryption methods, which are crucial to the current infrastructure of cyber security. Both RSA and ECDSA, the two most widely used security algorithms today, may be (in principle) solved by the Shor algorithm in polynomial time due to its ability to efficiently solve the discrete logarithm problem, potentially making present infrastructures insecure against a quantum attack. The National Institute of Standards and Technology (NIST) reacted with the post-quantum cryptography (PQC) standardization process to develop and optimize a series of post-quantum algorithms (PQAs) based on difficult mathematical problems that are not susceptible to being solved by Shor’s algorithm. Whilst high-powered computers can run these PQAs efficiently, further work is needed to investigate and benchmark the performance of these algorithms on lower-powered (constrained) devices and the ease with which they may be integrated into existing protocols such as TLS. This paper provides quantitative benchmark and handshake performance data for the most recently selected PQAs from NIST, tested on a Raspberry Pi 4 device to simulate today’s IoT (Internet of Things) devices, and provides quantitative comparisons with previous benchmarking data on a range of constrained systems. CRYSTALS-Kyber and CRYSTALS-Dilithium are shown to be the most efficient PQAs in the key encapsulation and signature algorithms, respectively, with Falcon providing the optimal TLS handshake size.

Multi-Modal in Vitro Experiments Mimicking the Flow Through a Mitral Heart Valve Phantom.

Fluid-structure interaction (FSI) models are more commonly applied in medical research as computational power is increasing. However, understanding the accuracy of FSI models is crucial, especially in the context of heart valve disease in patient-specific models. Therefore, this study aimed to create a multi-modal benchmarking data set for cardiac-inspired FSI models, based on clinically important parameters, such as the pressure, velocity, and valve opening, with an in vitro phantom setup. An in vitro setup was developed with a 3D-printed phantom mimicking the left heart, including a deforming mitral valve. A range of pulsatile flows were created with a computer-controlled motor-and-pump setup. Catheter pressure measurements, magnetic resonance imaging (MRI), and echocardiography (Echo) imaging were used to measure pressure and velocity in the domain. Furthermore, the valve opening was quantified based on cine MRI and Echo images. The experimental setup, with 0.5% cycle-to-cycle variation, was successfully built and six different flow cases were investigated. Higher velocity through the mitral valve was observed for increased cardiac output. The pressure difference across the valve also followed this trend. The flow in the phantom was qualitatively assessed by the velocity profile in the ventricle and by streamlines obtained from 4D phase-contrast MRI. A multi-modal set of data for validation of FSI models has been created, based on parameters relevant for diagnosis of heart valve disease. All data is publicly available for future development of computational heart valve models.

Experimental Design of Extractive Question-Answering Systems: Influence of Error Scores and Answer Length

Question-answering (QA) systems are becoming more and more important because they enable human-computer communication in a natural language. In recent years, significant progress has been made with transformer-based models that leverage deep learning in combination with large amounts of text data. However, a significant challenge with QA systems lies in their complexity rooted in the ambiguity and flexibility of a natural language. This makes even their evaluation a formidable task. For this reason, in this study, we focus on the evaluation of extractive question-answering (EQA) systems by conducting a large-scale analysis of distilBERT using benchmark data provided by the Stanford Question Answering Dataset (SQuAD). Specifically, the main objectives of this paper are fourfold. First, we study the influence of the answer length on the performance and we demonstrate that there is an inverse correlation between both. Second, we study differences in exact match (EM) measures because there are different definitions commonly used in the literature. As a result, we find that despite the fact that all of those measures are named ”exact match” these measures are actually different from each other. Third, we study the practical relevance of these different definitions because due to the ambivalent meaning of ”exact match” in the literature, it is often unclear if reported improvements are genuine or only due to a change in the exact match measure. Importantly, our results show that differences between differently defined EM measures are in the same order of magnitude as reported differences found in the literature. This raises concerns about the robustness of reported results. Fourth, we provide guidelines to improve the experimental design of general EQA studies, aiming to enhance performance evaluation and minimize the potential for spurious results.

ATOMIX benchmark datasets for dissipation rate measurements using shear probes

Turbulent mixing in the ocean, lakes and reservoirs facilitates the transport of momentum, heat, nutrients, and other passive tracers. Turbulent fluxes are proportional to the rate of turbulent kinetic energy dissipation per unit mass, ε. A common method for ε measurements is using microstructure profilers with shear probes. Such measurements are now widespread, and a non-expert practitioner will benefit from best practice guidelines and benchmark datasets. As a part of the Scientific Committee on Oceanographic Research (SCOR) working group on “Analysing ocean turbulence observations to quantify mixing” (ATOMIX), we compiled a collection of five benchmark data of ε from measurements of turbulence shear using shear probes. The datasets are processed using the ATOMIX recommendations for best practices documented separately. Here, we describe and validate the datasets. The benchmark collection is from different types of instruments and covers a wide range of environmental conditions. These datasets serve to guide the users to test their ε estimation methods and quality-assurance metrics, and to standardize their data for archiving.

Study on multi‐cycle gas‐water displacing mechanism in underground gas storage of low‐permeability reservoir based on PNM

AbstractA deep understanding of the pore‐scale multi‐cycle two‐phase seepage mechanism of gas‐water systems in low‐permeability reservoirs is crucial for enhancing oil and gas recovery and optimizing the operating conditions of underground gas storage. A pore network model was reconstructed using micro‐CT images of low‐permeability core samples from the Dagang Oilfield, China. The mathematical models of gas‐water flow in the pore‐scale models were established based on Poiseuille's law and the quasi‐static displacement theory, considering the gas compressibility and slip effects. The porosity, pore size, absolute permeability, and relative permeability (Kr) of oil‐water and gas‐water were calculated using pore network simulations and validated against experimental benchmark data of the same samples. The effects of multi‐cycle displacement, rock wettability, and average pore pressure on the gas‐water two‐phase flow were simulated and analyzed. The results showed that the relative permeability of gas (Krg) at the same water saturation level significantly increased when the gas compressibility and slip effects were considered. With an increase in the number of gas‐water displacement cycles, Krg at the same water saturation level showed a decreasing trend, indicating a reduction in the gas seepage capability. Krg decreased with increasing water contact angle. With an increase in the average pore pressure, Krg decreased and gradually increased when the average pressure exceeded 25 MPa.

Salzburg Intensive Care database (SICdb): a detailed exploration and comparative analysis with MIMIC-IV.

The utilization of artificial intelligence (AI) in healthcare is on the rise, demanding increased accessibility to (public) medical data for benchmarking. The digitization of healthcare in recent years has facilitated medical data scientists' access to extensive hospital data, fostering AI-based research. A notable addition to this trend is the Salzburg Intensive Care database (SICdb), made publicly available in early 2023. Covering over 27 thousand intensive care admissions at the University Hospital Salzburg from 2013 to 2021, this dataset presents a valuable resource for AI-driven investigations. This article explores the SICdb and conducts a comparative analysis with the widely recognized Medical Information Mart for Intensive Care - version IV (MIMIC-IV) database. The comparison focuses on key aspects, emphasizing the availability and granularity of data provided by the SICdb, particularly vital signs and laboratory measurements. The analysis demonstrates that the SICdb offers more detailed information with higher data availability and temporal resolution for signal data, especially for vital signs, compared to the MIMIC-IV. This is advantageous for longitudinal studies of patients' health conditions in the intensive care unit. The SICdb provides a valuable resource for medical data scientists and researchers. The database offers comprehensive and diverse healthcare data in a European country, making it well suited for benchmarking and enhancing AI-based healthcare research. The importance of ongoing efforts to expand and make public datasets available for advancing AI applications in the healthcare domain is emphasized by the findings.

Tensile testing data of additive manufactured ASTM D638 standard specimens with embedded internal geometrical features

Additive manufacturing (AM) is now widely used for research and industrial production. The benchmark data for mechanical properties of additively manufactured specimens is very useful for many communities. This data article presents a tensile testing dataset of ASTM D638 size specimens without and with embedded internal geometrical features printed using polylactic acid (PLA) in a Fused Filament Fabrication (FFF) additive manufacturing process. The added features can mimic defects of various shapes and sizes. This work is a supplement to the published research article Assisted defect detection by in-process monitoring of additive manufacturing using optical imaging and infrared thermography (Additive Manufacturing, 2023, 103483). The printed specimens were tensile tested. Stress-strain graphs were developed and used to calculate the mechanical properties such as ultimate tensile strength (UTS) and strain at UTS. The mechanical properties, the correlations between mechanical properties and size, shape and location of geometrical features (defects), and the trends in mechanical properties can be useful in benchmarking the results of other researchers.

Building Extraction from LiDAR Point Clouds Based on Revised RandLA-Net

Abstract. 3D building models is crucial for applications in smart cities. Automatic reconstruction of 3D buildings has been investigated based on various data sources. Point clouds from airborne LiDAR scanners can be used to extract buildings data due to its high accuracy and point density. In this paper, we present a methodology to segment buildings and corresponding rooftop structure from point clouds. First, RandLA-Net, which is an efficient and lightweight neural network for semantic segmentation of large-scale point clouds, is revised and adopted for building segmentation. By implementing local feature aggregation of each point, RandLA-Net can effectively preserve geometric details in point clouds. Besides 3D coordinates of point clouds, we incorporated point attributes including pulse intensity and return numbers into the network as additional features. Feature normalizations are applied to the input features. To achieve a better result of the local feature aggregation, hyperparameters of the network are fine-tuned according to the density of points and building size. Based on the classified building point clouds, DBSCAN clustering algorithm is implemented for segmenting individual buildings. Elevation histogram analysis is conducted to determine optimal threshold values for delineating candidate rooftop point clouds of individual buildings. For the buildings with multiple rooftops, multiple elevation threshold values are necessary to extract corresponding rooftops or walls. Then DBSCAN is employed again for segmentation of individual rooftops and denoising of point clouds of each building. Finally, Alpha-shape analysis is applied based on adaptive threshold values to build the envelope of each rooftop. Experiments show that our implementation of building segmentation using RandLA-net achieves higher mean IoU (Intersection over Union) and better classification performance in building segmentation. ISPRS benchmark data was used in our experiment and our methodology produce results with accuracy of 90.79%.

Imbalance-Robust Multi-Label Self-Adjusting kNN

In the task of multi-label classification in data streams, instances arriving in real time need to be associated with multiple labels simultaneously. Various methods based on the k Nearest Neighbors algorithm have been proposed to address this task. However, these methods face limitations when dealing with imbalanced data streams, a problem that has received limited attention in existing works. To approach this gap, this paper introduces the Imbalance-Robust Multi-Label Self-Adjusting kNN (IRMLSAkNN), designed to tackle multi-label imbalanced data streams. IRMLSAkNN’s strength relies on maintaining relevant instances with imbalance labels by using a discarding mechanism that considers the imbalance ratio per label. On the other hand, it evaluates subwindows with an imbalance-aware measure to discard older instances that are lacking performance. We conducted statistical experiments on 32 benchmark data streams, evaluating IRMLSAkNN against eight multi-label classification algorithms using common accuracy-aware and imbalance-aware measures. The obtained results demonstrate that IRMLSAkNN consistently outperforms these algorithms in terms of predictive capacity and time cost across various levels of imbalance.

PointGAT: A Quantum Chemical Property Prediction Model Integrating Graph Attention and 3D Geometry.

Predicting quantum chemical properties is a fundamental challenge for computational chemistry. While the development of graph neural networks has advanced molecular representation learning and property prediction, their performance could be further enhanced by incorporating three-dimensional (3D) structural geometry into two-dimensional (2D) molecular graph representation. In this study, we introduce the PointGAT model for quantum molecular property prediction, which integrates 3D molecular coordinates with graph-attention modeling. Comparison with other current models in molecular prediction tasks showed that PointGAT could provide higher predictive accuracy in various benchmark data sets from MoleculeNet, including ESOL, FreeSolv, Lipop, HIV, and 6 out of 12 tasks of the QM9 data set. To further examine PointGAT prediction of quantum mechanical (QM) energies, we constructed a C10 data set comprising 11,841 charged and chiral carbocation intermediates with QM energies calculated at the DM21/6-31G*//B3LYP/6-31G* levels. Notably, PointGAT achieved an R2 value of 0.950 and an MAE of 1.616 kcal/mol, outperforming even the best-performing graph neural network model with a reduction of 0.216 kcal/mol in MAE and an improvement of 0.050 in R2. Additional ablation studies indicated that incorporating molecular geometry into the model resulted in markedly higher predictive accuracy, reducing the MAE value from 1.802 to 1.616 kcal/mol. Moreover, visualization of PointGAT atomic attention weights suggested its predictions were interpretable. Findings in this study support the application of PointGAT as a powerful and versatile tool for quantum chemical property prediction that can facilitate high-accuracy modeling for fundamental exploration of chemical space as well as drug design and molecular engineering.

Light-Adaptive Human Body Key Point Detection Algorithm Based on Multi-Source Information Fusion.

The identification of key points in the human body is vital for sports rehabilitation, medical diagnosis, human-computer interaction, and related fields. Currently, depth cameras provide more precise depth information on these crucial points. However, human motion can lead to variations in the positions of these key points. While the Mediapipe algorithm demonstrates effective anti-shake capabilities for these points, its accuracy can be easily affected by changes in lighting conditions. To address these challenges, this study proposes an illumination-adaptive algorithm for detecting human key points through the fusion of multi-source information. By integrating key point data from the depth camera and Mediapipe, an illumination change model is established to simulate environmental lighting variations. Subsequently, the fitting function of the relationship between lighting conditions and adaptive weights is solved to achieve lighting adaptation for human key point detection. Experimental verification and similarity analysis with benchmark data yielded R2 results of 0.96 and 0.93, and cosine similarity results of 0.92 and 0.90. With a threshold range of 8, the joint accuracy rates for the two rehabilitation actions were found to be 89% and 88%. The experimental results demonstrate the stability of the proposed method in detecting key points in the human body under changing illumination conditions, its anti-shake ability for human movement, and its high detection accuracy. This method shows promise for applications in human-computer interaction, sports rehabilitation, and virtual reality.

Developing preliminary cost estimates for foundation systems of high-rise buildings

This study provided a unique Artificial intelligence (AI) model for the cost estimation of the foundation system for high-rise buildings. First, parametric data research was conducted, including 180 examples of foundation systems for high-rise building models related to criteria such as average column spacing, floor height, slap load, overall weight, base shear, and over. In addition, moment, gravity load, seismic load, design load, various columns (mid-edge-corner), soil profile, bearing capacity, GWT, excavation depth, and foundation systems are all factors to consider (shallow and deep footing). The research encompassed column spacing (5.0, 7.0, and 9.0 m) and the number of stories from 15 to 35, with an average cost decrease of 15%. The AI model utilizes extensive machine learning algorithms to analyze structural and geotechnical parameters. It provides a predictive tool that improves decision-making in the preliminary design. The methodology includes collecting data, training models utilizing innovative neural network techniques and validating them against current cost estimation models. The findings indicate that the AI model significantly improves the accuracy of cost estimations, with a precision rate of 95%, compared to traditional methods. Moreover, the model presents a potential cost savings of up to 15%. Construction designers and decision-makers may utilize the study’s findings to increase the accuracy of foundation system selection, lower costs and construction length, and improve overall project performance. Future work may involve applying this concept to infrastructure and substructure projects and collecting more data for benchmarking and empirical study.

MicroSimACC: an open database for field experiments on the potential capacity impact of commercial Adaptive Cruise Control (ACC)

Commercial availability of vehicle automation has become mainstream. Most of today’s new vehicles can perform longitudinal car following autonomously via Adaptive Cruise Control (ACC). Field experiments demonstrate that today’s commercially available ACC vehicles provide similar headways and capacities as human-driven vehicles on freeways under steady-state and free-flow conditions. However, field tests also demonstrated that the design of today’s commercially available ACC vehicles can lead to further capacity reduction when operating in non-steady-state conditions where queues are present and speeds frequently fluctuate. These experiments generated MicroSimACC, a comprehensive set of field data that encompasses full speed range car following with interruptions from lane change manoeuvres. This will benefit the research community by providing benchmark data for developing models to be integrated into microscopic simulations for more prospective analyses and planning.

A discrete Jaya algorithm for vehicle routing problems with uncertain demands

Vehicle Routing Problem with Uncertain Demands (VRPUD) is one of the research hotspots in the field of logistics scheduling. In this paper, a Discrete Jaya (DJaya) algorithm is presented for the VRPUD to minimize the total cost. A novel dividing-point-based coding scheme is designed to represent solutions with higher robustness. In addition, an efficient repair strategy is embedded into the decoding process to avoid the failure of producing feasible solutions. The best and worst solutions are employed to generate offspring solutions in DJaya. Several efficient local search methods are also presented to enhance the exploitation ability and increase the diversity of solutions. Based on the benchmark data sets of the VRPUD, numerical simulations are carried out for the proposed DJaya algorithm with uncertain demands. Computational results and comparisons with the state-of-the-art algorithms demonstrate the superiority of the proposed algorithm in solving VRPUD.

Assessment of Embedding Schemes in a Hybrid Machine Learning/Classical Potentials (ML/MM) Approach.

Machine learning (ML) methods have reached high accuracy levels for the prediction of in vacuo molecular properties. However, the simulation of large systems solely through ML methods (such as those based on neural network potentials) is still a challenge. In this context, one of the most promising frameworks for integrating ML schemes in the simulation of complex molecular systems are the so-called ML/MM methods. These multiscale approaches combine ML methods with classical force fields (MM), in the same spirit as the successful hybrid quantum mechanics-molecular mechanics methods (QM/MM). The key issue for such ML/MM methods is an adequate description of the coupling between the region of the system described by ML and the region described at the MM level. In the context of QM/MM schemes, the main ingredient of the interaction is electrostatic, and the state of the art is the so-called electrostatic-embedding. In this study, we analyze the quality of simpler mechanical embedding-based approaches, specifically focusing on their application within a ML/MM framework utilizing atomic partial charges derived in vacuo. Taking as reference electrostatic embedding calculations performed at a QM(DFT)/MM level, we explore different atomic charges schemes, as well as a polarization correction computed using atomic polarizabilites. Our benchmark data set comprises a set of about 80k small organic structures from the ANI-1x and ANI-2x databases, solvated in water. The results suggest that the minimal basis iterative stockholder (MBIS) atomic charges yield the best agreement with the reference coupling energy. Remarkable enhancements are achieved by including a simple polarization correction.

Evaluating Graph Layout Algorithms: A Systematic Review of Methods and Best Practices

AbstractEvaluations—encompassing computational evaluations, benchmarks and user studies—are essential tools for validating the performance and applicability of graph and network layout algorithms (also known as graph drawing). These evaluations not only offer significant insights into an algorithm's performance and capabilities, but also assist the reader in determining if the algorithm is suitable for a specific purpose, such as handling graphs with a high volume of nodes or dense graphs. Unfortunately, there is no standard approach for evaluating layout algorithms. Prior work holds a ‘Wild West’ of diverse benchmark datasets and data characteristics, as well as varied evaluation metrics and ways to report results. It is often difficult to compare layout algorithms without first implementing them and then running your own evaluation. In this systematic review, we delve into the myriad of methodologies employed to conduct evaluations—the utilized techniques, reported outcomes and the pros and cons of choosing one approach over another. Our examination extends beyond computational evaluations, encompassing user‐centric evaluations, thus presenting a comprehensive understanding of algorithm validation. This systematic review—and its accompanying website—guides readers through evaluation types, the types of results reported, and the available benchmark datasets and their data characteristics. Our objective is to provide a valuable resource for readers to understand and effectively apply various evaluation methods for graph layout algorithms. A free copy of this paper and all supplemental material is available at osf.io, and the categorized papers are accessible on our website at https://visdunneright.github.io/gd‐comp‐eval/.