Find research datasets worth reusing

Collu-Bench: A Benchmark for Predicting Language Model Hallucinations in Code

Despite their success, large language models (LLMs) face the critical challenge of hallucinations, generating plausible but incorrect content. While much research has focused on hallucinations in multiple modalities including images and natural language text, less attention has been given to hallucinations in source code, which leads to incorrect and vulnerable code that causes significant financial loss. To pave the way for research in LLMs' hallucinations in code, we introduce Collu-Bench, a benchmark for predicting code hallucinations of LLMs across code generation (CG) and automated program repair (APR) tasks. Collu-Bench includes 13,234 code hallucination instances collected from five datasets and 11 diverse LLMs, ranging from open-source models to commercial ones.  To better understand and predict code hallucinations, Collu-Bench provides detailed features such as the per-step log probabilities of LLMs' output, token types, and the execution feedback of LLMs' generated code for in-depth analysis. In addition, we conduct experiments to predict hallucination on Collu-Bench, using both traditional machine learning techniques and neural networks, which achieves 22.03 -- 33.15% accuracy. Our experiments draw insightful findings of code hallucination patterns, reveal the challenge of accurately localizing LLMs' hallucinations, and highlight the need for more sophisticated techniques.

Segmentation Dataset for Periorbital Segmentation and Distance Prediction

High quality segmentation of the eyes and lids is an essential step in developing clinically relevant deep learning models for oculoplastic and craniofacial surgery. However, there are currently no publicly available datasets suitable for this purpose. As such, we have developed and validated a novel dataset for oculoplastic segmentation and periorbital distance prediction. Using images from two open-source datasets, we segmented the iris, sclera, lid, caruncle, and brow from cropped eye images. Five trained annotators performed the segmentations, and intergrader reliability was assessed on 100 randomly selected images aftera two-week forgetting period, yielding an average Dice score of 0.82 ± 0.01. Intragrader reliability on 20 images averaged a Dice score of 0.81 ± 0.08. To demonstrate the dataset's utility, we trained three DeepLabV3 models following standard procedures. This first-of-its-kind dataset, along with a toolkit for periorbital distance prediction, is publicly available to support the development of clinically useful segmentation models for oculoplastic and craniofacial applications.

Machine learning reveals the diversity of human 3D chromatin contact patterns (example predictions genome wide)

Example data for the paper: Machine learning reveals the diversity of human 3D chromatin contact patterns GitHub: https://github.com/erin-n-gilbertson/3DGenome-diversity/tree/main biorXiv: https://www.biorxiv.org/content/10.1101/2023.12.22.573104v1.full Manuscript accepted at Molecular Biology and Evolution Of primary interest will be the example predictions genome wide for hg38 reference, human-archaic hominin ancestor and most divergent 1KG individual per genome along with the Jupyter notebook tutorial for making your own Akita predictions given any input 1MB sequence. <div> <ul> <li>bin: contains python script for and qsub array shell script for generating example predictions. These scripts can be modified to take in any fasta files as input.</li> <li>akita_predictions: contains both Akita prediction output arrays and SVG files with predicted contact maps for the hg38 reference, human-archaic hominin ancestor and most divergent 1KG individual in each of 4,873 1MB windows</li> <li>anc_window_spearman.csv: spearman correlation between each 1KG individual and the ancestor for each 1MB window. To calculate 3D divergence subtract these values from 1.</li> <li>basenji: basenji dir from their github, necessary in the directory to run predictions - https://github.com/calico/basenji/tree/master</li> <li>genomes: fasta genomes for hg38 reference and human-archaic hominin ancestor used to make akita predictions</li> <li>divergent_windows: variants and expected divergence distributions for 392 more divergent than expected windows. Defined in the manuscript as windows where 3D divergence between 1KG indiivudals and the ancestor is greater than what would be expected based on sequence divergence. See manuscript Fig. S9 for more details. </li> <li>windows.txt: 4,873 1MB genomic windows with 100% coverage in hg38 used for Akita predictions</li> <li>making_examples.ipynb: jupyter notebook with tutorial instructions for making Akita predictions on any human genome sequence.</li> </ul> </div>

Supplementary files: Machine Learning Insights into Türkiye's Climate Variability: Predictive Modelling and Spatial Analysis

This dataset and python code were used in the study titled "Machine Learning Insights into Türkiye's Climate Variability: Predictive Modelling and Spatial Analysis".

Reproducibility Case Study and Survey: Machine Learning-based Additive Manufacturing Process Monitoring and Quality Prediction

Machine learning (ML)-based monitoring systems have been extensively developed to enhance the print quality of additive manufacturing (AM). However, the reproducibility of the proposed ML-based AM monitoring systems in published works has not been investigated due to a lack of evaluation methods. In the paper 'Towards reproducible machine learning-based process monitoring and quality prediction research for additive manufacturing,' we propose a reproducibility investigation pipeline and conduct two case studies to validate the pipeline. This dataset records the data generated by one of the case studies. This dataset also contains the reproducibility survey results.

GPCR-BSD: Predicted Pockets data

Open the record for dataset details and reuse information.

openmit-licenseOct 2024View details →

Transitive prediction of small molecule function through alignment of high-content screening resources

This dataset supports the development of CLIPn, a contrastive-learning framework designed to align heterogeneous high-content screening (HCS) profile datasets. GitHub link: https://github.com/AltschulerWu-Lab/CLIPn <h2>Directory Structure</h2> <h3>Data Files</h3> <ul> <li>HCS_datasets.pkl: Contains 13 high-content screening (HCS) datasets from multiple studies across 20 years.</li> <li>Hypoxia.pkl: Contains 8 profile datasets using different assays and treated under diverse hypoxia durations.</li> <li>Expression.pkl: Contains 2 transcriptional profile datasets and 6 image profile datasets for multimodal analysis.</li> </ul> <h3>Folders</h3> raw_profiles: HCS13/ - Contains raw data from 13 high-content screening (HCS) datasets. Each dataset includes meta and feature files.  L1000/ - CDRP_feature_exp.csv: Raw L1000 expression data from the CDRP dataset. - CDRP_meta_exp.csv: Metadata associated with the CDRP expression data. - LINCS_feature_exp.csv: Raw L1000 expression data from the LINCS dataset. - LINCS_meta_exp.csv: Metadata associated with the LINCS expression data. RxRx3/ - RxRx3_feature_final.csv: Profile data from the RxRx3 dataset. - RxRx3_meta_final.csv: Metadata from the RxRx3 dataset. Uncharacterized_compounds/ - NCI_cpnData.csv: Feature data for uncharacterized compounds from the NCI dataset. - NCI_cpnInfo.csv: Information about uncharacterized compounds in the NCI dataset. - Prestwick_UTSW_cpnData.csv: Feature data for uncharacterized compounds from the Prestwick UTSW dataset. - Prestwick_UTSW_cpnInfo.csv: Information about uncharacterized compounds from the Prestwick UTSW dataset. <h2> Usage</h2> <code>import pickle</code> <code>with open('data.pkl', 'rb') as f:</code> <code>    data = pickle.load(f)</code> <code>X = data['X']</code> <code>y = data['y']</code> <h2>Data Reference</h2> For raw datasets from 13 HCS database, data and analysis pipeline for dataset 1 was obtained from https://www.science.org/doi/suppl/10.1126/science.1100709/suppl_file/perlman.som.zip; for datasets 2-3, data were shared by authors; For datasets 4-5, analysis code was downloaded from https://static-content.springer.com/esm/art%3A10.1038%2Fnbt.3419/MediaObjects/41587_2016_BFnbt3419_MOESM21_ESM.zip and data were shared by authors; For datasets 6-7, processed dataset was downloaded from AWS following instructions from https://github.com/carpenter-singh-lab/2022_Haghighi_NatureMethods, and replicate_level_cp_normalized.csv.gz features were used. For project datasets 8-13, datasets and analysis results were downloaded from https://zenodo.org/records/7352487. For RxRx3, dataset was obtained from https://www.rxrx.ai/rxrx3. L1000 transcript datasets were downloaded using the same link as datasets 6-7 and the processed transcript data files (named “replicate_level_l1k.csv”) were used.

Data for "A neural correlate of learning fails to predict foraging efficiency in the bumble bee Bombus terrestris"

This file contain the foraging and synaptic density data collected in the summer of 2019 by Pasquier, Pull, Ott and Leadbeater. Used in the paper titled "A neural correlate of learning fails to predict foraging efficiency in the bumble bee Bombus terrestris".

Supplementary Data for: Benchmark of density functional theory in the prediction of chemical shielding anisotropies for anisotropic NMR based structural elucidation

Additional  Data for the research paper titled: Benchmark of density functional theory in the prediction of chemical shielding anisotropies for anisotropic NMR based structural elucidation. Anisotropy Benchmark for Carbon NS372: <ul> <li>Chemical Shielding Tensor for the molecules in the NS372 test set for carbon (XLSX)</li> <li>Coordinate files for the molcules of the NS372 test set that contained carbon (in NS372-Carbon-COORD-Files.zip)</li> </ul> DFT Benchmark for RCSA for Natural Products: <ul> <li>Chemical Shielding Tensor used for the RCSA analysis of 6 Natural Products (CSV)</li> <li>Turbomole Input and Ouput files for the DFT calculation of the natural products (in RAW_DATA_for_RCSA_Analysis.zip)</li> <li>ConArch+ Input and Ouput files for the RCSA analysis using  (in RAW_DATA_for_RCSA_Analysis.zip)</li> <li>Coordinate files used for the RCSA analysis  (in RAW_DATA_for_RCSA_Analysis.zip)</li> </ul>

opencc-zeroAug 2024View details →

The data for Prediction of Large Solar Flares Based on SHARP and HED Magnetic Field Parameters

The repository includes HED and SHARP Datasets, along with descriptions of the data processing procedures, and the experimental code utilized in our study. The files named "HED parameters.csv" and "SHARP parameters.csv" represent the raw data of HED and SHARP Datasets, respectively. Each dataset consists of 286 ARs, specifically, each dataset includes 189 C-class flaring ARs, 86 M-class flaring ARs, and 11 X-class flaring ARs. The column name "AR" represents the NOAA AR numbers, the column name "time" represents the time at which the AR samples were acquired, and the column name "level" represents the flare class of the AR. Moreover, in the HED dataset, the column name "energy" represents the "E_{free}", "shear_mean" represents the "\Psi", "uj_mean" represents the "J_{Z}", "uhc_mean" represents the "H_{C}", "GBH_MEAN" represents the "B_{h}", and "ALP_MEAN" represents the "\alpha". In the SHARP dataset, the column names of the other columns correspond to the names of various parameters. The files named "SHARP_CV.zip" and "HED_CV.zip" represent the ten cross-validation sets that have been normalized and divided using an AR-based cross-validation method. The data processing procedures are as follows. Since the SHARP and HED parameters have distinct scales and units, they are individually normalized using mean-standard deviation. Subsequently, we set the labels for C-class flaring ARs to the negative class which are equal to 0, and the labels for M/X-class flaring ARs to positive class which are equal to 1. We employ an AR-based cross-validation (CV) method to partition the SHARP and HED datasets with the distribution of the training, validation, and testing sets at a ratio of 60%, 20%, and 20%, respectively. This process is repeated ten times, resulting in ten cross-validation sets in both SHARP datasets (SHARP_CV) and HED datasets (HED_CV), respectively. The column named "level" in the SHARP_CV data and the column named "key" in the HED_CV data represent the labels for the flare classes of the ARs. The file named "Parameter description and formula of SHARP and HED as well as data division process.pdf" represents the description and formula of SHARP and HED parameters, as well as the procedure for generating 10-fold cross-validation set divisions for the SHARP/HED datasets. We design five solar flare prediction models leveraging five currently popular deep learning algorithms: Transformer, BiLSTM-Attention, BiLSTM, LSTM-Attention, and LSTM. Moreover, we use the NN model as the baseline model to compare with other deep learning model. The solar flare prediction model used in this paper can be accessed via https://drive.google.com/drive/folders/1n6fXcQdCBogKt6L0aaPFagraFcXmX0fw?usp=drive_link.

Artifact of the paper: Scheduling with lightweight predictions in power-constrained HPC platforms

Please refer to the <a href="https://zenodo.org/records/13961003/files/artifact-overview.pdf?download=1&preview=1">artifact-overview.pdf</a> file in this dataset for instructions to reproduce the experiments we have conducted for this article, or for more context about the article.

opencc-by-4.0Jul 2024View details →

Mode-of-inheritance predictions for all possible missense variants in the human genome (hg38)

Ensemble and consensus approaches to prediction of recessive inheritance for missense variants in human disease.

opencc-by-4.0Apr 2024View details →

Dataset for article "Prediction of thermal shock induced cracking in multi-material ceramics using a stress-energy criterion" published in Engineering Fracture Mechanics

Dataset contains graphical outputs of the numerical analysis performed using Finite element method in finite elements software Ansys Mechanical. It contains also photoes of the tested specimens. Further, material data used for the numerical analysis and measured by authors are included in the csv file and all necessary input codes for the FE system Ansys creating data for graphs in the publication are provided in the subfolder "Models-Ansys" within "Data" directory.

opencc-by-4.0Sep 2024View details →

Machine Learning for Energy Consumption Prediction of Numerical Controlled Programs - NC Files

The NC files housed within this DOI represent the Data the Machine Learning Models were Trained/Validated/Tested on, during the execution of the work in the thesis, " Machine Learning for Energy Consumption Prediction of Numerically Controlled Programs." Theses files were created by Samuel D. Stencel, a Gradute Research Assistant at Purdue University.

opencc-by-4.0Nov 2024View details →

Temperature data corresponding to "Hybrid Phenology Modeling for Predicting Temperature Effects on Tree Dormancy"

MERRA2 Temperature data corresponding to "Hybrid Phenology Modeling for Predicting Temperature Effects on Tree Dormancy"

opencc-by-4.0Apr 2024View details →

Catchments do not strictly follow Budyko curves over multiple decades but deviations are minor and predictable

The supplementary data is provided in the form of an Excel file which consists of two sheets. ST-1 consists of data for individual catchments, and ST-2 consists of data for regional marginal distributions of deviations ST-1: The topographic characteristics for all study catchments (Latitude, Longitude, Area, Elevation), the length of the runoff data availability, mean precipitation for the period 1901-2015, mean 20-year evaporative indices for the five 20-year periods, median deviation values along with the skew normal distribution parameters and p-values of the Wilcoxon Signed Rank Test for each 20-year distribution deviation are provided. p-values of the Kolmogorov-Smirnov Test, performed on the two consecutive distribution of deviations, are provided in the next columns. Similarly, the long-term median deviation values, IQR of the deviations, fitted distribution parameters and temporal stability flag of aggregated marginal distribution of deviations along with associated changes in Q are provided. The last column shows the IQR of median deviation values for the 20 individual moving window aggregated marginal distribution of deviations. ST-2: Long-term median deviation values, IQR of deviations, and parameters of the fitted regional distributions of deviations along with the number of catchments in each category are provided.

opencc-zeroDec 2015View details →

Data from: Juvenile concentrations of IGF-1 predict life-history trade-offs in a wild mammal

Early postnatal development can have profound effects on life-history traits later in life. One mechanism hypothesized to mediate this relationship is the anabolic hormone, insulin-like growth factor-1 (IGF-1). IGF-1 contributes importantly to postnatal growth, and thus offers a means by which environmental and genetic variation might direct organismal development, reproduction and survival. We tested whether juvenile concentrations of IGF-1 can predict intraspecific variation in life-history traits later in life using longitudinal data from free-living female spotted hyenas (Crocuta crocuta). We found that juvenile concentrations of IGF-1 predicted heavier juvenile mass, which in turn predicted greater survival to reproductive maturity. However, independent of mass, higher juvenile concentrations of IGF-1 predicted earlier age at first parturition and reduced longevity in adulthood. Our results highlight the importance of early postnatal development as a determination period in mammals and suggest that concentrations of IGF-1 during this sensitive period can be used to predict important later-life trade-offs between growth, reproductive fitness and life span in wild, long-lived animals.

opencc-zeroOct 2019View details →

Data from: Antibiotic uptake across gram-negative outer membranes: better predictions towards better antibiotics

Crossing the gram-negative bacterial membrane poses a major barrier to antibiotic development, as many small molecules that can biochemically inhibit key bacterial processes are rendered microbiologically ineffective by their poor cellular uptake. The outer membrane is the major permeability barrier for many drug-like molecules, and the chemical properties that enable efficient uptake into mammalian cells fail to predict bacterial uptake. We have developed a computational method for accurate prospective prediction of outer-membrane uptake of drug-like molecules, which we combine with a new medium-throughput experimental assay of outer membrane vesicle swelling. Parallel molecular dynamics simulations of compound uptake through E. coli OmpF are used to successfully and quantitatively predict experimental permeabilities measured via either outer membrane swelling or via prior liposome swelling measurements. These simulations are analyzed using an inhomogeneous solubility-diffusion model to yield predictions of permeability. For most polar molecules we test, outer membrane permeability also correlates well with whole-cell uptake. The ability to accurately predict and measure outer-membrane uptake of a wide variety of small molecules will enable simpler determination of which molecular scaffolds and which derivatives are most promising prior to extensive chemical synthesis. It will also assist in formulating a more systematic understanding of the chemical determinants of outer-membrane permeability.

opencc-zeroDec 2017View details →

Data from: Local prey community composition and genetic distance predict venom divergence among populations of the northern Pacific rattlesnake (Crotalus oreganus)

Identifying the environmental correlates of divergence in functional traits between populations can provide insights into the evolutionary mechanisms that generate local adaptation. Here, we assess patterns of population differentiation in expressed venom proteins in Northern Pacific rattlesnakes (Crotalus oreganus) from 13 locations across California. We evaluate the relative importance of major biotic (prey species community composition), abiotic (temperature, precipitation, and elevation) and genetic factors (genetic distance based on RADseq loci) as correlates of population divergence in venom phenotypes. We found that over half of the variation in venom composition is associated with among-population differentiation for genetic and environmental variables, and that this variation occurred along axes defining previously observed functional trade-offs between venom proteins that have neurotoxic, myotoxic and hemorrhagic effects. Surprisingly, genetic differentiation among populations was the best predictor of venom divergence, accounting for 46% of overall variation, whereas differences in prey community composition and abiotic factors explained smaller amounts of variation (23% and 19%, respectively). The association between genetic differentiation and venom composition could be due to an isolation by distance effect or, more likely it may reflect an isolation-by-environment effect where selection against recent migrants is strong, producing a correlation between neutral genetic differentiation and venom differentiation. Our findings suggest that even coarse estimates of prey community composition can be useful in understanding the selection pressures acting on patterns of venom protein expression. Additionally, our results suggest that factors other than adaptation to spatial variation in prey need to be considered when explaining population divergence in venom.