Dataset, splits, models, and scripts from the manuscript "When Do Quantum Mechanical Descriptors Help Graph Neural Networks Predict Chemical Properties?" are provided. The curated dataset includes 37 QM descriptors for 64,921 unique molecules across six levels of theory: wB97XD, B3LYP, M06-2X, PBE0, TPSS, and BP86. This dataset is stored in the data.tar.gz file, which also contains a file for multitask constraints applied to various atomic and bond properties. The data splits (training, validation, and test splits) for both random and scaffold-based divisions are saved as separate index files in splits.tar.gz. The trained D-MPNN models for predicting QM descriptors are saved in the models.tar.gz file. The scripts.tar.gz file contains ready-to-use scripts for training machine learning models to predict QM descriptors, as well as scripts for predicting QM descriptors using our trained models on unseen molecules and for applying radial basis function (RBF) expansion to QM atom and bond features.

Below are descriptions of the available scripts:

1. atom_bond_descriptors.sh: Trains atom/bond targets.
2. atom_bond_descriptors_predict.sh: Predicts atom/bond targets from pre-trained model.
3. dipole_quadrupole_moments.sh: Trains dipole and quadrupole moments.
4. dipole_quadrupole_moments_predict.sh: Predicts dipole and quadrupole moments from pre-trained model.
5. energy_gaps_IP_EA.sh: Trains energy gaps, ionization potential (IP), and electron affinity (EA).
6. energy_gaps_IP_EA_predict.sh: Predicts energy gaps, IP, and EA from pre-trained model.
7. get_constraints.py: Generates constraints file for testing dataset. This generated file needs to be provided before using our trained models to predict the atom/bond QM descriptors of your testing data.
8. csv2pkl.py: Converts QM atom and bond features to .pkl files using RBF expansion for use with Chemprop software.

Below is the procedure for running the ml-QM-GNN on your own dataset:

1. Use get_constraints.py to generate a constraint file required for predicting atom/bond QM descriptors with the trained ML models.
2. Execute atom_bond_descriptors_predict.sh to predict atom and bond properties. Run dipole_quadrupole_moments_predict.sh and energy_gaps_IP_EA_predict.sh to calculate molecular QM descriptors.
3. Utilize csv2pkl.py to convert the data from predicted atom/bond descriptors .csv file into separate atom and bond feature files (which are saved as .pkl files here).
4. Run Chemprop to train your models using the additional predicted features supported here.

BackgroundThis study proposes machine learning-driven data preparation (MLDP) for optimal data preparation (DP) prior to building prediction models for cancer cohorts.MethodsA collection of well-established DP methods were incorporated for building the DP pipelines for various clinical cohorts prior to machine learning. Evolutionary algorithm principles combined with hyperparameter optimization were employed to iteratively select the best fitting subset of data preparation algorithms for the given dataset. The proposed method was validated for glioma and prostate single center cohorts by 100-fold Monte Carlo (MC) cross-validation scheme with 80-20% training-validation split ratio. In addition, a dual-center diffuse large B-cell lymphoma (DLBCL) cohort was utilized with Center 1 as training and Center 2 as independent validation datasets to predict cohort-specific clinical endpoints. Five machine learning (ML) classifiers were employed for building prediction models across all analyzed cohorts. Predictive performance was estimated by confusion matrix analytics over the validation sets of each cohort. The performance of each model with and without MLDP, as well as with manually-defined DP were compared in each of the four cohorts.ResultsSixteen of twenty established predictive models demonstrated area under the receiver operator characteristics curve (AUC) performance increase utilizing the MLDP. The MLDP resulted in the highest performance increase for random forest (RF) (+0.16 AUC) and support vector machine (SVM) (+0.13 AUC) model schemes for predicting 36-months survival in the glioma cohort. Single center cohorts resulted in complex (6-7 DP steps) DP pipelines, with a high occurrence of outlier detection, feature selection and synthetic majority oversampling technique (SMOTE). In contrast, the optimal DP pipeline for the dual-center DLBCL cohort only included outlier detection and SMOTE DP steps.ConclusionsThis study demonstrates that data preparation prior to ML prediction model building in cancer cohorts shall be ML-driven itself, yielding optimal prediction models in both single and multi-centric settings.

Context

The two datasets are related to red and white variants of the Portuguese "Vinho Verde" wine. For more details, consult the reference [Cortez et al., 2009]. Due to privacy and logistic issues, only physicochemical (inputs) and sensory (the output) variables are available (e.g. there is no data about grape types, wine brand, wine selling price, etc.).

These datasets can be viewed as classification or regression tasks. The classes are ordered and not balanced (e.g. there are much more normal wines than excellent or poor ones).

This dataset is also available from the UCI machine learning repository, https://archive.ics.uci.edu/ml/datasets/wine+quality , I just shared it to kaggle for convenience. (If I am mistaken and the public license type disallowed me from doing so, I will take this down if requested.)

Content

For more information, read [Cortez et al., 2009].
Input variables (based on physicochemical tests):
1 - fixed acidity
2 - volatile acidity
3 - citric acid
4 - residual sugar
5 - chlorides
6 - free sulfur dioxide
7 - total sulfur dioxide
8 - density
9 - pH
10 - sulphates
11 - alcohol
Output variable (based on sensory data):
12 - quality (score between 0 and 10)

Tips

What might be an interesting thing to do, is aside from using regression modelling, is to set an arbitrary cutoff for your dependent variable (wine quality) at e.g. 7 or higher getting classified as 'good/1' and the remainder as 'not good/0'. This allows you to practice with hyper parameter tuning on e.g. decision tree algorithms looking at the ROC curve and the AUC value. Without doing any kind of feature engineering or overfitting you should be able to get an AUC of .88 (without even using random forest algorithm)

KNIME is a great tool (GUI) that can be used for this.
1 - File Reader (for csv) to linear correlation node and to interactive histogram for basic EDA.
2- File Reader to 'Rule Engine Node' to turn the 10 point scale to dichtome variable (good wine and rest), the code to put in the rule engine is something like this:
- $quality$ > 6.5 => "good"
- TRUE => "bad"
3- Rule Engine Node output to input of Column Filter node to filter out your original 10point feature (this prevent leaking)
4- Column Filter Node output to input of Partitioning Node (your standard train/tes split, e.g. 75%/25%, choose 'random' or 'stratified')
5- Partitioning Node train data split output to input of Train data split to input Decision Tree Learner node and
6- Partitioning Node test data split output to input Decision Tree predictor Node
7- Decision Tree learner Node output to input Decision Tree Node input
8- Decision Tree output to input ROC Node.. (here you can evaluate your model base on AUC value)

Inspiration

Use machine learning to determine which physiochemical properties make a wine 'good'!

Acknowledgements

This dataset is also available from the UCI machine learning repository, https://archive.ics.uci.edu/ml/datasets/wine+quality , I just shared it to kaggle for convenience. (I am mistaken and the public license type disallowed me from doing so, I will take this down at first request. I am not the owner of this dataset.

Please include this citation if you plan to use this database: P. Cortez, A. Cerdeira, F. Almeida, T. Matos and J. Reis. Modeling wine preferences by data mining from physicochemical properties. In Decision Support Systems, Elsevier, 47(4):547-553, 2009.

Relevant publication

P. Cortez, A. Cerdeira, F. Almeida, T. Matos and J. Reis. Modeling wine preferences by data mining from physicochemical properties. In Decision Support Systems, Elsevier, 47(4):547-553, 2009.

CIFAR-100

The CIFAR-10 and CIFAR-100 dataset contains labeled subsets of the 80 million tiny images dataset. They were collected by Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. * More info on CIFAR-100: https://www.cs.toronto.edu/~kriz/cifar.html * TensorFlow listing of the dataset: https://www.tensorflow.org/datasets/catalog/cifar100 * GitHub repo for converting CIFAR-100 tarball files to png format: https://github.com/knjcode/cifar2png

All images were sized 32x32 in the original dataset

The CIFAR-10 dataset consists of 60,000 32x32 colour images in 10 classes, with 6,000 images per class. There are 50,000 training images and 10,000 test images [in the original dataset].

This dataset is just like the CIFAR-10, except it has 100 classes containing 600 images each. There are 500 training images and 100 testing images per class. The 100 classes in the CIFAR-100 are grouped into 20 superclasses. Each image comes with a "fine" label (the class to which it belongs) and a "coarse" label (the superclass to which it belongs). However, this project does not contain the superclasses. * Superclasses version: https://universe.roboflow.com/popular-benchmarks/cifar100-with-superclasses/

More background on the dataset: https://i.imgur.com/5w8A0Vm.png" alt="CIFAR-100 Dataset Classes and Superclassees">

Version 1 (original-images_Original-CIFAR100-Splits):

• Original images, with the original splits for CIFAR-100: train (83.33% of images - 50,000 images) set and test (16.67% of images - 10,000 images) set only.
• This version was not trained

Version 2 (original-images_trainSetSplitBy80_20):

• Original, raw images, with the train set split to provide 80% of its images to the training set (approximately 40,000 images) and 20% of its images to the validation set (approximately 10,000 images)
• Trained from Roboflow Classification Model's ImageNet training checkpoint
• https://blog.roboflow.com/train-test-split/ https://i.imgur.com/kSPeKGn.png" alt="Train/Valid/Test Split Rebalancing">

Citation:

@TECHREPORT{Krizhevsky09learningmultiple,
  author = {Alex Krizhevsky},
  title = {Learning multiple layers of features from tiny images},
  institution = {},
  year = {2009}
}

Overview

The data has been split into two groups:

• training set (train.csv)
• test set (test.csv)

The training set should be used to build your machine learning models. For the training set, we provide the outcome (also known as the “ground truth”) for each passenger. Your model will be based on “features” like passengers’ gender and class. You can also use feature engineering to create new features.

The test set should be used to see how well your model performs on unseen data. For the test set, we do not provide the ground truth for each passenger. It is your job to predict these outcomes. For each passenger in the test set, use the model you trained to predict whether or not they survived the sinking of the Titanic.

We also include gender_submission.csv, a set of predictions that assume all and only female passengers survive, as an example of what a submission file should look like.

Data Dictionary:

| Variable | Definition | Key | | --- | --- | | survival | Survival | 0 = No, 1 = Yes | | pclass | Ticket class | 1 = 1st, 2 = 2nd, 3 = 3rd | | sex | Sex | | | Age | Age in years | | | sibsp | # of siblings / spouses aboard the Titanic | | | parch | # of parents / children aboard the Titanic | | | ticket | Ticket number | | | fare | Passenger fare | | | cabin | Cabin number | | | embarked | Port of Embarkation | C = Cherbourg, Q = Queenstown, S = Southampton |

Variable Notes

pclass: A proxy for socio-economic status (SES) 1st = Upper 2nd = Middle 3rd = Lower

age: Age is fractional if less than 1. If the age is estimated, is it in the form of xx.5

sibsp: The dataset defines family relations in this way... Sibling = brother, sister, stepbrother, stepsister Spouse = husband, wife (mistresses and fiancés were ignored)

parch: The dataset defines family relations in this way... Parent = mother, father Child = daughter, son, stepdaughter, stepson Some children travelled only with a nanny, therefore parch=0 for them.

More - Find More Exciting🙀 Datasets Here - An Upvote👍 A Dayᕙ(`▿´)ᕗ , Keeps Aman Hurray Hurray..... ٩(˘◡˘)۶Hehe

The CIFAR-10 dataset is a large dataset of labeled images. It is a popular dataset for machine learning and artificial intelligence research. The dataset consists of 60,000 32x32 images. These images are split into 10 mutually exclusive classes, with 6,000 images per class. The classes are airplanes, automobiles, birds, cats, deer, dogs, frogs, horses, ships, and trucks.

Offline reinforcement learning (RL) is a promising direction that allows RL agents to be pre-trained from large datasets avoiding recurrence of expensive data collection. To advance the field, it is crucial to generate large-scale datasets. Compositional RL is particularly appealing for generating such large datasets, since 1) it permits creating many tasks from few components, and 2) the task structure may enable trained agents to solve new tasks by combining relevant learned components. This submission provides four offline RL datasets for simulated robotic manipulation created using the 256 tasks from CompoSuite Mendez et al., 2022. In every task in CompoSuite, a robot arm is used to manipulate an object to achieve an objective all while trying to avoid an obstacle. There are for components for each of these four axes that can be combined arbitrarily leading to a total of 256 tasks. The component choices are * Robot: IIWA, Jaco, Kinova3, Panda* Object: Hollow box, box, dumbbell, plate* Objective: Push, pick and place, put in shelf, put in trashcan* Obstacle: None, wall between robot and object, wall between goal and object, door between goal and object The four included datasets are collected using separate agents each trained to a different degree of performance, and each dataset consists of 256 million transitions. The degrees of performance are expert data, medium data, warmstart data and replay data: * Expert dataset: Transitions from an expert agent that was trained to achieve 90% success on every task.* Medium dataset: Transitions from a medium agent that was trained to achieve 30% success on every task.* Warmstart dataset: Transitions from a Soft-actor critic agent trained for a fixed duration of one million steps.* Medium-replay-subsampled dataset: Transitions that were stored during the training of a medium agent up to 30% success. These datasets are intended for the combined study of compositional generalization and offline reinforcement learning. Methods The datasets were collected by using several deep reinforcement learning agents trained to the various degrees of performance described above on the CompoSuite benchmark (https://github.com/Lifelong-ML/CompoSuite) which builds on top of robosuite (https://github.com/ARISE-Initiative/robosuite) and uses the MuJoCo simulator (https://github.com/deepmind/mujoco). During reinforcement learning training, we stored the data that was collected by each agent in a separate buffer for post-processing. Then, after training, to collect the expert and medium dataset, we run the trained agents for 2000 trajectories of length 500 online in the CompoSuite benchmark and store the trajectories. These add up to a total of 1 million state-transitions tuples per dataset, totalling a full 256 million datapoints per dataset. The warmstart and medium-replay-subsampled dataset contain trajectories from the stored training buffer of the SAC agent trained for a fixed duration and the medium agent respectively. For medium-replay-subsampled data, we uniformly sample trajectories from the training buffer until we reach more than 1 million transitions. Since some of the tasks have termination conditions, some of these trajectories are trunctated and not of length 500. This sometimes results in a number of sampled transitions larger than 1 million. Therefore, after sub-sampling, we artificially truncate the last trajectory and place a timeout at the final position. This can in some rare cases lead to one incorrect trajectory if the datasets are used for finite horizon experimentation. However, this truncation is required to ensure consistent dataset sizes, easy data readability and compatibility with other standard code implementations. The four datasets are split into four tar.gz folders each yielding a total of 12 compressed folders. Every sub-folder contains all the tasks for one of the four robot arms for that dataset. In other words, every tar.gz folder contains a total of 64 tasks using the same robot arm and four tar.gz files form a full dataset. This is done to enable people to only download a part of the dataset in case they do not need all 256 tasks. For every task, the data is separately stored in an hdf5 file allowing for the usage of arbitrary task combinations and mixing of data qualities across the four datasets. Every task is contained in a folder that is named after the CompoSuite elements it uses. In other words, every task is represented as a folder named

Dataset Card

This is a dataset containing ML ArXiv papers. The dataset is a version of the original one from CShorten, which is a part of the ArXiv papers dataset from Kaggle. Three steps are made to process the source data:

useless columns removal; train-test split; ' ' removal and trimming spaces on sides of the text.

Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms

Authors:

• Han Xiao, Kashif Rasul and Roland Vollgraf
• https://arxiv.org/abs/1708.07747

Dataset Obtained From: https://github.com/zalandoresearch/fashion-mnist

All images were sized 28x28 in the original dataset

Fashion-MNIST is a dataset of Zalando's article images—consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. We intend Fashion-MNIST to serve as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms. It shares the same image size and structure of training and testing splits. * Source

Here's an example of how the data looks (each class takes three-rows): https://github.com/zalandoresearch/fashion-mnist/raw/master/doc/img/fashion-mnist-sprite.png" alt="Visualized Fashion MNIST dataset">

Version 1 (original-images_Original-FashionMNIST-Splits):

• Original images, with the original splits for MNIST: train (86% of images - 60,000 images) set and test (14% of images - 10,000 images) set only.
• This version was not trained

Version 3 (original-images_trainSetSplitBy80_20):

• Original, raw images, with the train set split to provide 80% of its images to the training set and 20% of its images to the validation set
• https://blog.roboflow.com/train-test-split/ https://i.imgur.com/angfheJ.png" alt="Train/Valid/Test Split Rebalancing">

Citation:

@online{xiao2017/online,
 author    = {Han Xiao and Kashif Rasul and Roland Vollgraf},
 title    = {Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms},
 date     = {2017-08-28},
 year     = {2017},
 eprintclass = {cs.LG},
 eprinttype  = {arXiv},
 eprint    = {cs.LG/1708.07747},
}

This dataset corresponds to the paper "BETH Dataset: Real Cybersecurity Data for Anomaly Detection Research" by Kate Highnam* (@jinxmirror13), Kai Arulkumaran* (@kaixhin), Zachary Hanif*, and Nicholas R. Jennings (@LboroVC).

This paper was published in the ICML Workshop on Uncertainty and Robustness in Deep Learning 2021 and Conference on Applied Machine Learning for Information Security (CAMLIS 2021)

THIS DATASET IS STILL BEING UPDATED

Context

When deploying machine learning (ML) models in the real world, anomalous data points and shifts in the data distribution are inevitable. From a cyber security perspective, these anomalies and dataset shifts are driven by both defensive and adversarial advancement. To withstand the cost of critical system failure, the development of robust models is therefore key to the performance, protection, and longevity of deployed defensive systems.

We present the BPF-extended tracking honeypot (BETH) dataset as the first cybersecurity dataset for uncertainty and robustness benchmarking. Collected using a novel honeypot tracking system, our dataset has the following properties that make it attractive for the development of robust ML methods: 1. At over eight million data points, this is one of the largest cyber security datasets available 2. It contains modern host activity and attacks 3. It is fully labelled 4. It contains highly structured but heterogeneous features 5. Each host contains benign activity and at most a single attack, which is ideal for behavioural analysis and other research tasks. In addition to the described dataset

Further data is currently being collected and analysed to add alternative attack vectors to the dataset.

There are several existing cyber security datasets used in ML research, including the KDD Cup 1999 Data (Hettich & Bay, 1999), the 1998 DARPA Intrusion Detection Evaluation Dataset (Labs, 1998; Lippmann et al., 2000), the ISCX IDS 2012 dataset (Shiravi et al., 2012), and NSL-KDD (Tavallaee et al., 2009), which primarily removes duplicates from the KDD Cup 1999 Data. Each includes millions of records of realistic activity for enterprise applications, with labels for attacks or benign activity. The KDD1999, NSLKDD, and ISCX datasets contain network traffic, while the DARPA1998 dataset also includes limited process calls. However, these datasets are at best almost a decade old, and are collected on in-premise servers. In contrast, BETH contains modern host activity and activity collected from cloud services, making it relevant for current real-world deployments. In addition, some datasets include artificial user activity (Shiravi et al., 2012) while BETH contains only real activity. BETH is also one of the few datasets to include both kernel-process and network logs, providing a holistic view of malicious behaviour.

Content

The BETH dataset currently represents 8,004,918 events collected over 23 honeypots, running for about five noncontiguous hours on a major cloud provider. For benchmarking and discussion, we selected the initial subset of the process logs. This subset was further divided into training, validation, and testing sets with a rough 60/20/20 split based on host, quantity of logs generated, and the activity logged—only the test set includes an attack

The dataset is composed of two sensor logs: kernel-level process calls and network traffic. The initial benchmark subset only includes process logs. Each process call consists of 14 raw features and 2 hand-crafted labels.

See the paper for more details. For details on the events recorded within the logs, see this report.

Benchmarks

Code for our benchmarks, as detailed in the paper, are available through Github at: https://github.com/jinxmirror13/BETH_Dataset_Analysis

Acknowledgements

Thank you to Dr. Arinbjörn Kolbeinsson for his assistance in analysing the data and the reviewers for their positive feedback.

Semi-flexible docking was performed using AutoDock Vina 1.2.2 software on the SARS-CoV-2 main protease M^pro (PDB ID: 6WQF).

Two data sets are provided in the xyz format containing the AutoDock Vina docking scores. These files were used as input and/or reference in the machine learning models using TensorFlow, XGBoost, and SchNetPack to study their docking scores prediction capability. The first data set originally contained 60,411 in-vivo labeled compounds selected for the training of ML models. The second data set,denoted as in-vitro-only, originally contained 175,696 compounds active or assumed to be active at 10 μM or less in a direct binding assay. These sets were downloaded on the 10th of December 2021 from the ZINC15 database. Four compounds in the in-vivo set and 12 in the in-vitro-only set were left out of consideration due to presence of Si atoms. Compounds with no charges assigned in mol2 files were excluded as well (523 compounds in the in-vivo and 1,666 in the in-vitro-only set). Gasteiger charges were reassigned to the remaining compounds using OpenBabel. In addition, four in-vitro-only compounds with docking scores greater than 1 kcal/mol have been rejected.

The provided in-vivo and the in-vitro-only sets contain 59,884 (in-vivo.xyz) and 174,014 (in-vitro-only.xyz) compounds, respectively. Compounds in both sets contain the following elements: H, C, N, O, F, P, S, Cl, Br, and I. The in-vivo compound set was used as the primary data set for the training of the ML models in the referencing study.

The file in-vivo-splits-data.csv contains the exact composition of all (random) 80-5-15 train-validation-test splits used in the study, labeled I, II, III, IV, and V. Eight additional random subsets in each of the in-vivo 80-5-15 splits were created to monitor the training process convergence. These subsets were constructed in such a manner, that each subset contains all compounds from the previous subset (starting with the 10-5-15 subset) and was enlarged by one eighth of the entire (80-5-15) train set of a given split. These subsets are further referred to as in_vivo_10_(I, II, ..., V), in_vivo_20_(I, II, ..., V),..., in_vivo_80_(I, II, ... V).

Hate Speech and Offensive Language Detection

Hate Speech and Offensive Language Detection on Twitter

By hate_speech_offensive (From Huggingface) [source]

About this dataset

This dataset, named hate_speech_offensive, is a meticulously curated collection of annotated tweets with the specific purpose of detecting hate speech and offensive language. The dataset primarily consists of English tweets and is designed to train machine learning models or algorithms in the task of hate speech detection. It should be noted that the dataset has not been divided into multiple subsets, and only the train split is currently available for use.

The dataset includes several columns that provide valuable information for understanding each tweet's classification. The column count represents the total number of annotations provided for each tweet, whereas hate_speech_count signifies how many annotations classified a particular tweet as hate speech. On the other hand, offensive_language_count indicates the number of annotations categorizing a tweet as containing offensive language. Additionally, neither_count denotes how many annotations identified a tweet as neither hate speech nor offensive language.

For researchers and developers aiming to create effective models or algorithms capable of detecting hate speech and offensive language on Twitter, this comprehensive dataset offers a rich resource for training and evaluation purposes

How to use the dataset

• Introduction:

• Dataset Overview:

  • The dataset is presented in a CSV file format named 'train.csv'.
  • It consists of annotated tweets with information about their classification as hate speech, offensive language, or neither.
  • Each row represents a tweet along with the corresponding annotations provided by multiple annotators.
  • The main columns that will be essential for your analysis are: count (total number of annotations), hate_speech_count (number of annotations classifying a tweet as hate speech), offensive_language_count (number of annotations classifying a tweet as offensive language), neither_count (number of annotations classifying a tweet as neither hate speech nor offensive language).

• Data Collection Methodology: The data collection methodology used to create this dataset involved obtaining tweets from Twitter's public API using specific search terms related to hate speech and offensive language. These tweets were then manually labeled by multiple annotators who reviewed them for classification purposes.

• Data Quality: Although efforts have been made to ensure the accuracy of the data, it is important to acknowledge that annotations are subjective opinions provided by individual annotators. As such, there may be variations in classifications between annotators.

• Preprocessing Techniques: Prior to training machine learning models or algorithms on this dataset, it is recommended to apply standard preprocessing techniques such as removing URLs, usernames/handles, special characters/punctuation marks, stop words removal, tokenization, stemming/lemmatization etc., depending on your analysis requirements.

• Exploratory Data Analysis (EDA): Conducting EDA on the dataset will help you gain insights and understand the underlying patterns in hate speech and offensive language. Some potential analysis ideas include:

  • Distribution of tweet counts per classification category (hate speech, offensive language, neither).
  • Most common words/phrases associated with each class.
  • Co-occurrence analysis to identify correlations between hate speech and offensive language.

• Building Machine Learning Models: To train models for automatic detection of hate speech and offensive language, you can follow these steps: a) Split the dataset into training and testing sets for model evaluation purposes. b) Choose appropriate features/

Research Ideas

• Sentiment Analysis: This dataset can be used to train models for sentiment analysis on Twitter data. By classifying tweets as hate speech, offensive language, or neither, the dataset can help in understanding the sentiment behind different tweets and identifying patterns of negative or offensive language.
• Hate Speech Detection: The dataset can be used to develop models that automatically detect hate speech on Twitter. By training machine learning algorithms on this annotated dataset, it becomes possible to create systems that can identify and flag hate speech in real-time, making social media platforms safer and more inclusive.
• Content Moderation: Social media platforms can use this dataset to improve their content m...

Boosted top tagging is an essential binary classification task for experiments at the Large Hadron Collider (LHC) to measure the properties of the top quark. The ATLAS Top Tagging Open Data Set is a publicly available dataset for the development of Machine Learning (ML) based boosted top tagging algorithms. The dataset consists of a nominal piece used for the training and evaluation of algorithms, and a systematic piece used for estimating the size of systematic uncertainties produced by an algorithm. The nominal data are is split into two orthogonal sets, named train and test. The systematic varied data is split into many more pieces that should only be used for evaluation in most cases. Both nominal sets are composed of equal parts signal (jets initiated by a boosted top quark) and background (jets initiated by light quarks or gluons).

A brief overview of these datasets is as follows. For more detailed information see arxiv:2047.20127.

• train_nominal - 92,820,427 jets, equal parts signal and background
• test_nominal - 10,306,813 jets, equal parts signal and background
• esup - 10,032,472 jets with the cluster energy scale up systematic variation active, equal parts signal and background
• esdown - 10,032,472 jets with the cluster energy scale down systematic variation active, equal parts signal and background
• cer - 10,040,653 jets with the cluster energy resolution systematic variation active, equal parts signal and background
• cpos - 10,032,472 jets with the cluster energy position systematic variation active, equal parts signal and background
• teg - 7,421,204 jets with the track efficiency global systematic variation active, 30% signal jets
• tej - 7,017,046 jets with the track efficiency in jets systematic variation active, 32% signal jets
• tfl - 5,907,310 jets with the track fake rate loose systematic variation active, 18% signal jets
• tfj - 6,977,371 jets with the track fake rate in jets systematic variation active, 32% signal jets
• bias - 10,011,330 jets with the track bias systematic variation active, 52% signal jets
• ttbar_pythia - 193,792 jets from Pythia simulated events containing Standard Model top-anti top quark pair production, all signal jets
• ttbar_herwig - 180,811 jets from Herwig simulated events containing Standard Model top-anti top quark pair production, all signal jets
• cluster - 5,000,004 jets simulated using the Sherpa cluster based hadronization model, all background jets
• string - 5,000,001 jets simulated using the Lund string based hadronization model, all background jets
• angular - 4,900,000 jets simulated using the Herwig angular ordered parton shower model, all background jets
• dipole - 4,900,000 jets simulated using the Herwig dipole parton shower model, all background jets

For each jet, the datasets contain:

• The four vectors of constituent particles
• 15 high level summary quantities evaluated on the jet
• The four vector of the whole jet
• A training weight (nominal only)
• PYTHIA shower weights (nominal only)
• A signal (1) vs background (0) label

There are two rules for using this data set: the contribution to a loss function from any jet should always be weighted by the training weight, and any performance claim is incomplete without an estimate of the systematic uncertainties via the method illustrated in this repository. The ideal model shows high performance but also small systematic uncertainties.

Overview

This replication package accompanies the paper "How Do Machine Learning Models Change?" In this study, we conducted a comprehensive analysis of over 200,000 commits and 1,200 releases across more than 50,000 models on the Hugging Face (HF) platform. Our goal was to understand how machine learning (ML) models evolve over time by classifying commit types based on an extended ML change taxonomy and analyzing patterns in commit and release activities using Bayesian networks.

Our research addresses three main aspects:

1. Categorization of Commit Changes: We classified over 200,000 commits on HF using an extended ML change taxonomy, providing a detailed breakdown of change types and their distribution across models.
2. Analysis of Commit Sequences: We examined the sequence and dependencies of commit types using Bayesian networks to identify temporal patterns and common progression paths in model changes.
3. Release Analysis: We investigated the distribution and evolution of release types, analyzing how model attributes and metadata change across successive releases.

This replication package contains all the necessary code, datasets, and documentation to reproduce the results presented in the paper.

Data Collection and Preprocessing

Data Collection

We collected data from the Hugging Face platform using the Hugging Face Hub API and the `HfApi` class. The data extraction was performed on November 6th, 2023. The collected data includes:

• Model Information: Details of over 380,000 models, including dataset sizes, training hardware, evaluation metrics, model file sizes, number of downloads and likes, tags, and the raw text of model cards.
• Commit Histories: Comprehensive commit details, including commit messages, dates, authors, and the list of files edited in each commit.
• Release Information: Information on model releases marked by tags in their repositories.

To enrich the commit data with detailed file change information, we integrated the PyDriller framework within the HFCommunity dataset.

Data Preprocessing

Commit Diffs

We computed the differences between commits for key files, specifically JSON configuration files (e.g., `config.json`). For each commit that modifies these files, we compared the changes with the previous commit affecting the same file to identify added, deleted, and updated keys.

Commit Classification

We classified each commit according to Bhatia et al.'s ML change taxonomy using the Gemini 1.5 Flash Large Language Model (LLM). This classification, using LLMs to apply Bhatia et al.'s taxonomy on a large-scale ML repository, is one of the main contributions of our paper. We ensured the correctness of the classification by achieving a Cohen's kappa coefficient ≥ 0.9 through iterative validation. In addition, we performed classification based on Swanson's categories using a simpler neural network approach, following methods from prior work. This classification has less impact compared to the detailed classification using Bhatia et al.'s taxonomy.

Model Metadata

We extracted detailed metadata from the model files of selected releases, focusing on attributes such as the number of parameters, tensor shapes, etc. We also calculated the differences between the metadata of successive releases.

Folder Structure

The replication package is organized as follows:

- code/: Contains the Jupyter notebooks with the data extraction, preprocessing, analysis, and model training scripts.

• Collection/: Contains two Jupyter notebooks for data collection:
  • HFTotalExtraction.ipynb: Script for collecting data on the entire Hugging Face platform.
  • HFReleasesExtraction.ipynb: Script for collecting data on models that contain releases.
• Preprocessing/: Contains preprocessing scripts:
  • HFTotalPreprocessing.ipynb: Preprocesses the dataset obtained from `HFTotalExtraction.ipynb`.
  • HFCommitsPreprocessing.ipynb: Processes commit data, including:
    • Retrieval of diff information between commits.
    • Classification of commits following Bhatia et al.'s taxonomy using LLMs.
    • Extension and adaptation of the final commits dataset, including additional variables for Bayesian network analysis.
  • HFReleasesPreprocessing.ipynb: Processes release data, including classification and preparation for analysis.
• Analysis/: Contains three Jupyter notebooks with the analysis for each research question:
  • RQ1_Analysis.ipynb: Analysis for RQ1.
  • RQ2_Analysis.ipynb: Analysis for RQ2.
  • RQ3_Analysis.ipynb: Analysis for RQ3.

- datasets/: Contains the raw, processed, and manually curated datasets used for the analysis.

• Main Datasets:
  • HFCommits_50K_RANDOM.csv: Contains the commits of 50,000 randomly sampled models from HF with the classification based on Bhatia et al.'s taxonomy.
  • HFCommits_MultipleCommits.csv: Contains the commits of 10,000 models with at least 10 commits, used for analyzing commit sequences.
  • HFReleases.csv: Contains over 1,200 releases from 127 models, classified using Bhatia et al.'s taxonomy.
  • model_metadata_with_diff.csv: Contains the metadata of releases from 27 models, including differences between successive releases.
  • These datasets correspond to the following dataset splits:
    • +200,000 commits from 50,000 models: Used for RQ1. Provides a broad overview of commit types and patterns across diverse models.
    • +200,000 commits from 10,000 models: Used for RQ2. Focuses on models with at least 10 commits for detailed evolutionary study.
    • +1,200 releases from 127 models: Used for RQ3.1, RQ3.2, and RQ3.3. Facilitates the investigation of release patterns and their evolution.
    • Metadata of 173 releases from 27 models: Used for RQ3.4. Analyzes the evolution of model parameters and configurations.

• Additional Datasets:
  • HF_Total_Raw.csv: Contains a snapshot of the entire Hugging Face platform with over 380,000 models, as obtained from HFTotalExtraction.ipynb.
  • HF_Total_Preprocessed.csv: Contains the preprocessed version of the entire HF dataset, as obtained from HFTotalPreprocessing.ipynb. This dataset is needed for the commits preprocessing.
  • Auxiliary datasets generated during processing are also included to facilitate reproduction of specific parts of the code without time-consuming steps.

- metadata/: Contains the tags_metadata.yaml file used during preprocessing.

- models/: Contains the model trained to classify commit messages into corrective, perfective, and adaptive types based on Swanson's traditional software maintenance categories.

- requirements.txt: Lists the required Python packages to set up the environment and run the code.

Setup and Execution

Prerequisites

• Python 3.10.11 or later.
• Jupyter Notebook or JupyterLab.

Installation

1. Download and Extract the Replication Package
2. Create a Virtual Environment (Recommended):bash
python -m venv venv
source venv/bin/activate # On Windows, use venv\Scripts\activate
3. Install Required Packages:bash
pip install -r requirements.txt

Notes

- LLM Usage: The classification of commits using the Gemini 1.5 Flash LLM requires access to the model. Ensure you have the necessary permissions and API keys to use the model.

- Computational Resources: Processing large datasets and running Bayesian network analyses may require significant computational resources. It is recommended to use a machine with ample memory and processing power.

- Reproducing Results: The auxiliary datasets included can be used to reproduce specific parts of the code without re-running the entire data collection and preprocessing pipeline.

Additional Information

Contact: If you have any questions or encounter issues, please contact the authors at joel.castano@upc.edu.

This README provides detailed instructions and information to reproduce and understand the analyses performed in the paper. If you find this package useful, please cite our work.

The Chinese Speech Separation Dataset (CSSD) is a dataset of audio recordings of people speaking Mandarin Chinese in a variety of noisy environments. The dataset consists of 10,000 audio recordings, each of which is a mixture of two speakers. The dataset is split into a training set of 8,000 recordings and a test set of 2,000 recordings. The audio recordings are in .wav format and have a sampling rate of 16 kHz. The audio recordings are labeled with the identities of the two speakers in the mixture. The CSSD dataset is a valuable resource for training speech separation models.

This dataset contains product prices from Amazon USA, with a focus on price prediction. With a good amount of data on what price points sell the most, you can train machine learning models to predict the optimal price for a product based on its features and product name.

If you find this dataset useful, make sure to show your appreciation by upvoting! ❤️✨

Inspirations

This dataset is a superset of my Amazon USA product price dataset. Another inspiration is this competition that awareded 100K Prize Money

What To Do?

• Your objective is to create a prediction model that will assist sellers in pricing their products within the optimal price range to generate the most sales.
• The dataset includes various data points, such as the number of reviews, rating, best seller status, and items sold last month.
• You can select specific factors (e.g., over 100 reviews = optimal price for the product) and then divide the dataset into products priced optimally vs products priced unoptimally.
• By utilizing techniques like vectorizing product names and features, you can train a model to provide the optimal price for a product, which sellers or businesses might find valuable.

How to know if a product sells?

• I would prefer to use the number of reviews as a metric to determine if a product sells. More reviews = more sales, right?
• According to one source only 1-2% of buyers leave a review
• So if we multiply the reviews for a product by 50x, then we would get a good understanding how many units has sold.
• If we then multiple the product price by number of units sold, we'd get the total revenue generated by the product

How is this useful?

• Sellers and businesses can leverage your model to determine the optimal price for their products, thereby maximizing sales.
• Businesses can assess the profitability of a product and plan their supply chain accordingly.

Smoking vs Non-Smoking Image Dataset: A collection of images labeled by smoking habits (Smoking vs Non-Smoking), pre-split into Training, Validation, and Test sets.

Machine learning (ML) of quantum mechanical properties shows promise for accelerating chemical discovery. For transition metal chemistry where accurate calculations are computationally costly and available training data sets are small, the molecular representation becomes a critical ingredient in ML model predictive accuracy. We introduce a series of revised autocorrelation functions (RACs) that encode relationships of the heuristic atomic properties (e.g., size, connectivity, and electronegativity) on a molecular graph. We alter the starting point, scope, and nature of the quantities evaluated in standard ACs to make these RACs amenable to inorganic chemistry. On an organic molecule set, we first demonstrate superior standard AC performance to other presently available topological descriptors for ML model training, with mean unsigned errors (MUEs) for atomization energies on set-aside test molecules as low as 6 kcal/mol. For inorganic chemistry, our RACs yield 1 kcal/mol ML MUEs on set-aside test molecules in spin-state splitting in comparison to 15–20× higher errors for feature sets that encode whole-molecule structural information. Systematic feature selection methods including univariate filtering, recursive feature elimination, and direct optimization (e.g., random forest and LASSO) are compared. Random-forest- or LASSO-selected subsets 4–5× smaller than the full RAC set produce sub- to 1 kcal/mol spin-splitting MUEs, with good transferability to metal–ligand bond length prediction (0.004–5 Å MUE) and redox potential on a smaller data set (0.2–0.3 eV MUE). Evaluation of feature selection results across property sets reveals the relative importance of local, electronic descriptors (e.g., electronegativity, atomic number) in spin-splitting and distal, steric effects in redox potential and bond lengths.

Dataset split for training and testing.

The activation and transformation of methane have long posed significant challenges in scientific research. The quest for highly active species and a profound understanding of the mechanisms of methane activation are pivotal for the rational design of related catalysts. In this study, by assembling a data set encompassing a total of 134 gas-phase metal species documented in the literature for methane activation via the mechanism of oxidative addition, machine learning (ML) models based on the backpropagation artificial neural network algorithm have been established with a range of intrinsic electronic properties of these species as features and the experimental rate constants of the reactions with methane as the target variables. It turned out that the satisfactory ML models could be described in terms of four key features, including the vertical electron detachment energy (VDE), the absolute value of the energy gap between the highest occupied molecular orbital of CH4, and the lowest unoccupied molecular orbital of the metal species (|ΔEH′–L|), the maximum natural charge of metal atoms (Qmax), and the maximum electron occupancy of valence s orbitals on metal atoms (ns_max), based on the feature selection complemented with manual intervention. The stability and generalization ability of the constructed model was validated using a specially designed data-splitting strategy and newly incorporated data. This study proved the feasibility and discussed the limitations of the ML model, which is described by four key features to predict the reactivity of metal-containing species toward methane through oxidative addition mechanisms. Furthermore, a careful preparation of the training data set that covers the full expected range of target and feature values aiming to achieve good predictive accuracy is suggested as a practical guideline for future research.