AMSGradBiasCorrection |
Which bias corrections AMSGrad applies to its moment estimates. |
AbmeTrainingStage |
ABME’s training stage (Park et al. |
AcquisitionFunctionType |
Represents different types of acquisition functions used in Bayesian optimization. |
ActivationCategory |
Categories of activation functions based on their architectural role and behavior. |
ActivationFunction |
Represents different activation functions used in neural networks and deep learning. |
ActivationFunctionRole |
Defines the functional roles of activation functions in neural networks. |
ActivationTask |
Tasks or architectural positions where an activation function is commonly used. |
AdamAnomalyGuardMode |
Policy for the PyTorch GradScaler-style anomaly guard on the Adam optimizer’s tape-based Step. |
AdditiveDecompositionAlgorithmType |
Represents different algorithm types for additive decomposition of time series data. |
AdversarialAttackMethod |
An adversarial attack that robustness testing can run against a trained model. |
AdversarialNormType |
The norm that bounds an adversarial perturbation or a robustness certificate. |
AdversarialPreprocessingMethod |
A built-in input-transformation defense that removes adversarial perturbations before the model sees the input. |
AnomalyDetectorType |
Specifies the type of anomaly detection algorithm to use. |
AssignmentStrategy |
Strategy for assigning requests to model versions during A/B testing. |
AttentionKernelPath |
Identifies which attention kernel a layer’s most recent forward pass actually executed. |
AttentionVariant |
Specifies the attention variant that determines how Key/Value heads map to Query heads. |
AutoAttackVersion |
The attack ensemble that AutoAttack runs (Croce and Hein 2020). |
AutoMLBudgetPreset |
Defines compute budget presets for AutoML runs. |
AutoMLFinalModelSelectionPolicy |
Defines how AutoML chooses the final model to return after the search completes. |
AutoMLSearchStrategy |
Defines the search strategy used to explore AutoML candidate configurations. |
AutoMLStatus |
Represents the current status of an AutoML search process. |
AutoMLTaskFamily |
Defines the high-level task family for an AutoML run. |
BenchmarkExecutionStatus |
Represents the execution outcome for a benchmark suite run. |
BenchmarkFailurePolicy |
Controls how the benchmark runner should behave when one or more suites fail. |
BenchmarkMetric |
Defines standardized metrics used in benchmark reports. |
BenchmarkReportDetailLevel |
Specifies how much detail should be included in benchmark reports. |
BenchmarkSuite |
Defines the supported benchmark suites available through the AiDotNet facade. |
BenchmarkSuiteKind |
Categorizes benchmark suites by their evaluation style and infrastructure requirements. |
BetaSchedule |
Defines the types of beta (noise variance) schedules available for diffusion models. |
BeveridgeNelsonAlgorithmType |
Represents different algorithm types for Beveridge-Nelson decomposition of time series data. |
BiasMode |
Whether a layer that can carry an additive bias should actually do so. |
BidiagonalAlgorithmType |
Represents different algorithm types for bidiagonal matrix decomposition. |
BoundaryHandlingMethod |
Specifies how to handle boundaries when processing data that extends beyond the available range. |
BoxAreaConvention |
How a box’s width and height are measured from its (x1, y1, x2, y2) corners when computing areas and IoU. |
CLIPVariant |
Variants for CLIP text encoder models. |
CacheEvictionPolicy |
Cache eviction policies for managing limited cache memory. |
CalibrationMethod |
Calibration methods for quantization - techniques to determine optimal scaling factors when converting high-precision models to low-precision formats. |
CausalDiscoveryAlgorithmType |
Specifies the algorithm to use for causal structure learning (DAG discovery). |
CausalDiscoveryCategory |
Categories of causal discovery algorithms based on their methodology. |
CertifiedDefenseMethod |
The method that certifies a prediction against bounded input perturbations. |
ChainType |
Represents different types of chains for composing language model operations. |
ChatClientRecordingMode |
Says whether chat calls are made live, saved to a store, or served back from one. |
ChatGLM3Variant |
Variants for the ChatGLM3 text encoder. |
CheckpointMetricType |
Standard metrics for checkpoint selection and early stopping. |
CholeskyAlgorithmType |
Represents different algorithm types for Cholesky decomposition of matrices. |
ClassWeightingStrategy |
How a class-weighted feature selector turns class frequencies into class weights. |
ClassificationTaskType |
Specifies the type of classification task being performed. |
ClusteringMetricType |
Defines the types of metrics used to evaluate clustering quality. |
ComponentType |
Defines the type of an AI pipeline component (Tier 2 metadata). |
CompressionType |
Defines the types of model compression strategies available in the AiDotNet library. |
ComputationalBudget |
Describes the computational budget available for model training and search. |
ComputeCost |
Relative computational cost of an operation. |
ConcertoModelSize |
Defines the backbone size variants for Concerto (joint 2D-3D point cloud segmentation). |
ConditionNumberMethod |
Specifies different methods for calculating the condition number of a matrix. |
ConformalPredictionMode |
Defines conformal prediction calibration modes. |
ContrastiveLossType |
Types of contrastive loss functions for knowledge distillation. |
ConvexSolverType |
Types of convex solvers available for MetaOptNet. |
CreditRule |
Selects the built-in credit-assignment (“learning”) rule used to produce per-parameter updates during neural-network training via AiModelBuilder.ConfigureCreditRule(...). |
CrossValidationType |
Defines the types of cross-validation strategies available. |
DEVAModelSize |
Defines the backbone size variants for DEVA (Decoupled Video Segmentation). |
DataComplexity |
Represents the level of complexity in a dataset, which helps determine appropriate model selection and preprocessing. |
DataSetType |
Represents the different types of datasets used in machine learning workflows. |
DecompositionComponentType |
Represents the different components that can be extracted when decomposing a time series. |
DeepARScaleHandling |
Selects how DeepAR rescales a series before the recurrent trunk sees it. |
DeformableModulationNormalization |
How a deformable convolution normalizes its modulation scalars. |
DenseNetVariant |
Specifies the DenseNet model variant. |
DepthAnythingV2ModelSize |
Defines the encoder size variants for Depth Anything V2. |
DiffusionPredictionType |
Defines what the diffusion model predicts during the denoising process. |
DiffusionReconstructionLoss |
The distance a diffusion model that predicts the clean sample x_0 trains with. |
DiffusionTrainingSampleSpace |
What a latent diffusion model’s training sample is: an image its VAE encodes, or a latent already encoded. |
DistanceMetricType |
Represents different methods for measuring the distance or similarity between data points. |
DistilledT5Variant |
Variants for distilled T5 text encoder models. |
DistributedStrategy |
Defines the distributed training strategy to use. |
DistributionType |
Represents different probability distributions used in statistical modeling and machine learning. |
EASTGeometryType |
The shape EAST predicts for each detected piece of text. |
EMDAlgorithmType |
Represents different algorithm types for Empirical Mode Decomposition (EMD). |
EdgeDeviceType |
Types of edge devices for optimization targeting. |
EdgeDirection |
Specifies the direction of edges to retrieve when querying a knowledge graph. |
EfficientNetVariant |
Specifies the EfficientNet model variant. |
EfficientTAMModelSize |
Defines the backbone size variants for EfficientTAM (Track Anything Model). |
EigenAlgorithmType |
Represents different algorithm types for computing eigenvalues and eigenvectors of matrices. |
EmbeddingPoolingStrategy |
How a text embedding model collapses a sequence of token representations into a single vector for the whole piece of text. |
EnvelopeType |
Specifies whether to use an upper or lower envelope in signal processing and data analysis operations. |
EoMTModelSize |
Defines the backbone size variants for EoMT (Encoder-only Mask Transformer). |
EvolveBlockStatus |
Reports how well a program source matched the configured evolve-block markers. |
ExpressionNodeType |
Defines the different types of nodes that can exist in a computational graph. |
FeatureExtractionStrategy |
Defines strategies for extracting features from higher-dimensional tensors. |
FedGenGeneratorOptimizer |
Update rule used to train the FedGen server generator. |
FedGenTeacherLoss |
Selects the teacher term of the FedGen server generator objective (Zhu et al. |
FederatedPartitioningStrategy |
Defines how a centralized dataset should be partitioned into per-client datasets for federated simulations. |
FederatedScoreAggregation |
How a federated feature selector combines the per-client feature scores into one global score. |
FencedCodeSelectionSource |
Records which fallback level produced the code returned by a fenced-code extraction. |
FewShotSelectionStrategy |
Represents strategies for selecting few-shot examples in prompt templates. |
FitType |
Represents different types of model fit quality and common issues in machine learning models. |
FitnessCalculatorType |
Specifies different loss functions and fitness calculators for evaluating model performance. |
FluxPredictorVariant |
Variants for FLUX noise predictor architectures. |
FluxVariant |
Variants for FLUX model. |
FoundationModelSize |
Defines the size variants available for foundation models. |
FourKAgentRestorePreference |
4KAgent’s restoration preference (the profile’s RestorePerference), which decides the super-resolution experts the agent may choose from. |
FunASRNanoTrainingStage |
Which modules a Fun-ASR-nano training step updates: the report’s supervised fine-tuning stages. |
GemmaVariant |
Variants for Gemma text encoder models. |
GeneticNodeType |
Types of nodes in a genetic programming tree. |
GradientType |
Specifies different types of gradient descent optimization algorithms used in machine learning. |
GramSchmidtAlgorithmType |
Represents different algorithm types for the Gram-Schmidt orthogonalization process. |
GraniteSpeechTrainingStage |
What a Granite-speech training step optimizes (Saon et al. |
GraphGenerationType |
Type of graph generation approach. |
GraphPooling |
Defines how a graph model combines its per-node representations into a single representation for the whole graph. |
GraphRAGMode |
Specifies the retrieval mode for enhanced GraphRAG. |
GuidanceType |
Types of guidance methods for diffusion model inference. |
HessenbergAlgorithmType |
Represents different algorithm types for Hessenberg decomposition of matrices. |
HiDreamVariant |
Variants for HiDream-I1 model. |
HodrickPrescottAlgorithmType |
Represents different algorithm types for implementing the Hodrick-Prescott filter. |
ImportanceThresholdStrategy |
Defines strategies for setting the importance threshold in feature selection. |
InfraType |
Defines the type of infrastructure component (Tier 3 metadata). |
InitializationMethod |
Methods for initializing a population. |
InputType |
Specifies the dimensionality of input data for machine learning models. |
InternImageModelSize |
Defines the model size variants for InternImage (DCNv3-based CNN backbone). |
Interpolation2DType |
Specifies different methods for interpolating 2D data points to create a continuous surface. |
InverseType |
Specifies different algorithms for calculating matrix inverses in mathematical operations. |
KGEmbeddingType |
Specifies the type of knowledge graph embedding model to use. |
KMaXDeepLabModelSize |
Defines the backbone size variants for kMaX-DeepLab panoptic segmentation. |
KandinskyVersion |
Versions for the Kandinsky model family. |
KernelType |
Specifies different kernel functions used in machine learning algorithms like Support Vector Machines (SVMs). |
LLMProvider |
Defines the large language model (LLM) providers available for AI agent assistance during model building and inference. |
LanguageModelBackbone |
Defines the language model backbone types used in multimodal neural networks. |
LanguageModelTrainableScope |
Which parts of a multimodal language model a training step may update. |
LayerApiShape |
Describes the Forward method signature shape a layer uses. |
LayerTask |
Tasks or architectural roles that a neural network layer performs. |
LayerType |
Specifies different types of layers used in neural networks, particularly in deep learning models. |
LdlAlgorithmType |
Represents different algorithm types for LDL decomposition of matrices. |
LicenseKeyStatus |
Represents the current status of a license key after validation. |
LinkPredictionDecoder |
Defines how a link-prediction model scores a candidate edge from the two node embeddings at its ends. |
LossApiShape |
Describes the method signature shape a loss function uses for its primary calculation. |
LossCategory |
Categories of loss functions based on the type of learning task they serve. |
LossTask |
Specific tasks a loss function is designed or commonly used for. |
LossTestInputFormat |
Describes what kind of test data a loss function expects. |
LossType |
Defines different loss functions used to measure how well a model’s predictions match the actual values. |
LqAlgorithmType |
Represents different algorithm types for LQ decomposition of matrices. |
LuAlgorithmType |
Represents different algorithm types for LU decomposition of matrices. |
MMDiTXVariant |
Variants for MMDiT noise predictor architectures. |
Mask2FormerModelSize |
Defines the backbone size variants for Mask2Former. |
MaskDINOModelSize |
Defines the backbone size variants for Mask DINO. |
MatrixDecompositionType |
Specifies different methods for breaking down (decomposing) matrices into simpler components. |
MatrixLayout |
Specifies how data is organized in matrices when working with arrays of data. |
MatrixType |
Defines the different types of matrices that can be used in mathematical operations. |
MaxFeatureSelection |
How many features a tree considers at each split, for the random-forest family. |
MbPAOutputDistribution |
Which likelihood MbPA’s output head models, i.e. |
MedGANDataType |
The kind of variables medGAN’s autoencoder is being fitted to. |
MedNeXtModelSize |
Defines the backbone size variants for MedNeXt medical segmentation. |
MedSAMModelSize |
Defines the backbone size variants for MedSAM (Medical SAM). |
MedicalAsrDecoderType |
Which end-to-end decoder a medical-conversation ASR model uses. |
MetricOptimizationDirection |
Specifies the direction for metric optimization (whether lower or higher values are better). |
MetricType |
Defines the types of metrics used to evaluate machine learning models. |
MixedPrecisionType |
Types of mixed precision training data types. |
MixedQueryTransformerModelSize |
Defines the backbone size variants for MixedQueryTransformer (MQ-Former). |
MixtureOfExpertsGatingNoise |
How a MixtureOfExpertsLayer scales the Gaussian noise it adds to the gating logits during training. |
MobileNetV2WidthMultiplier |
Specifies the width multiplier for MobileNetV2. |
MobileNetV3Variant |
Specifies the MobileNetV3 model variant. |
MobileNetV3WidthMultiplier |
Specifies the width multiplier for MobileNetV3. |
ModelCategory |
Defines the algorithm family or category that a machine learning model belongs to. |
ModelComplexity |
Indicates the computational complexity and resource requirements of a machine learning model. |
ModelCompressionMode |
Defines the mode of model compression to apply during serialization. |
ModelDomain |
Defines the application domain or field that a machine learning model is designed for. |
ModelPerformance |
Represents the overall performance quality of a machine learning model. |
ModelTask |
Defines the specific task or capability that a machine learning model performs. |
ModificationType |
Represents the types of modifications that can be applied to a model structure. |
MultiVectorAggregation |
How a multi-vector retriever combines the query similarities of one document’s vectors into a single score. |
MultiplicativeAlgorithmType |
Represents different multiplicative algorithm types for time series analysis and forecasting. |
NERModelVariant |
Defines common model size variants for Named Entity Recognition (NER) models. |
NetworkComplexity |
Defines the complexity level of a neural network architecture. |
NeuralArchitectureSearchStrategy |
Specifies the strategy used for neural architecture search. |
NeuralNetworkTaskType |
Defines the different types of tasks that a neural network can be designed to perform. |
NnUNetModelSize |
Defines the architecture variants for nnU-Net v2 medical segmentation. |
NormalizationMethod |
Defines different methods for normalizing data before processing in machine learning algorithms. |
NumericalReproducibility |
How strictly a layer must reproduce its defined output when a faster route is available. |
ODISEModelSize |
Defines the backbone size variants for ODISE (Open-vocabulary DIffusion-based panoptic SEgmentation). |
OMGSegModelSize |
Defines the backbone size variants for OMG-Seg. |
ObliviousTreeChoiceFunction |
The functions an ObliviousDecisionTreeLayer uses to choose a feature per level and to make each soft split decision. |
OneFormerModelSize |
Defines the backbone size variants for OneFormer. |
OperationType |
Represents different operation types in computation graphs for JIT compilation and automatic differentiation. |
OptimizationMode |
Specifies the mode of optimization for an optimizer. |
OptimizationPassType |
Represents the type of optimization pass applied to the computation graph. |
OptimizerKind |
Identifies the optimizer a research paper specifies for a model. |
OptimizerMomentStorage |
How an Adam-family optimizer stores its first and second moments (m, v): the bulk of a large model’s training memory after the weights themselves. |
OptimizerType |
Defines different optimization algorithms used to train machine learning models. |
OutlierDetectionMethod |
Defines different methods for detecting outliers in datasets. |
OutputDistribution |
Specifies the target distribution for quantile transformation. |
OutputType |
The expected format of model outputs that a loss function operates on. |
PIDNetModelSize |
Defines the backbone size variants for PIDNet real-time segmentation. |
PNAAggregator |
Aggregation function types for Principal Neighbourhood Aggregation (PNA). |
PNAScaler |
Scaler function types for Principal Neighbourhood Aggregation (PNA). |
ParameterType |
Defines the types of parameters that can be used in hyperparameter search |
PartitionStrategy |
Strategies for partitioning models between cloud and edge devices. |
PatchGANReceptiveField |
The discriminator receptive-field sizes reported by pix2pix (Isola et al., 2017), section 6.1.2. |
PayloadEncryptionScheme |
Specifies the encryption scheme applied to the model payload within an AIMF envelope. |
Phi4AudioTrainingStage |
What a Phi-4-multimodal speech training step updates (Abdin et al. |
PiiType |
A kind of personally identifiable information the text safety pipeline can detect. |
PipelineStage |
Defines which stage of an AI pipeline a component operates in. |
PixArtVariant |
Variants for the PixArt model family. |
PointSerializationOrder |
The built-in space-filling curve orders Point Transformer V3 serializes points with. |
PointTransformerV3Revision |
The two published Point Transformer V3 backbones in Pointcept. |
PolarAlgorithmType |
Represents different algorithm types for computing the polar decomposition of matrices. |
PoolingType |
Defines different methods for pooling (downsampling) data in neural networks, particularly in convolutional neural networks. |
PositionalEncodingType |
Represents the type of positional encoding used in transformer attention layers. |
PrecisionMode |
Defines the numeric precision mode for neural network training and computation. |
PredictionType |
Specifies the type of prediction task that a machine learning model performs. |
PreferenceLossType |
Types of loss functions for preference optimization methods. |
ProbabilityCalibrationMethod |
Defines probability calibration strategies for classification-like outputs. |
ProgramArtifactTier |
Says where an artifact store physically keeps one artifact’s bytes. |
ProgramDiffFailureReason |
Names why one SEARCH/REPLACE block could not be parsed or applied. |
ProgramEvolutionMode |
Selects how a language model is asked to change a candidate program. |
ProgramMetricAggregationStrategy |
Selects the rule that turns a dictionary of evaluator metrics into one scalar quality. |
ProgramMetricIssueReason |
Explains why one metric did not contribute to a scalar quality in the way a caller might expect. |
ProgramMetricValueKind |
Distinguishes the three kinds of value an evaluator may report under one metric name. |
ProgramNoveltyStage |
Names the stage of the novelty pipeline that settled a candidate’s fate. |
ProgramNoveltyVerdict |
A novelty judge’s answer about one candidate, including the case where it did not answer. |
ProgramOutputComparison |
Selects how a candidate program’s captured output is compared with an expected output. |
ProgramPromptEvolutionMode |
Says which request a prompt makes of the model: targeted edits, a full rewrite, or either by size. |
ProgramPromptExampleKind |
Says why an example program was chosen for a prompt, which decides how it is labelled. |
ProgramPromptFragmentKey |
Names one short, overridable phrase that prompt sections are assembled from. |
ProgramPromptSystemMessageMode |
Says how a configured system message is interpreted: as a template name or as literal text. |
ProgramPromptTemplateKey |
Names one overridable prompt template used when a language model evolves a program. |
ProgramProposalOutcome |
Records what happened to one request sent to a language model while proposing a candidate program. |
ProgramRunOutputTrigger |
Says what caused a best-program snapshot to be written to disk. |
ProgramSandboxMode |
Selects where a candidate program is executed when a program-evolution run scores it. |
PromptOptimizationStrategy |
Represents strategies for optimizing prompts to improve language model performance. |
PromptTemplateType |
Represents different types of prompt templates for language model interactions. |
QATMethod |
Specifies the Quantization-Aware Training (QAT) method to use during model training. |
QrAlgorithmType |
Represents different algorithm types for computing the QR decomposition of matrices. |
QualityLevel |
Quality levels for adaptive inference on resource-constrained devices. |
QuantizationGranularity |
Specifies the granularity level for quantization scaling factors. |
QuantizationMode |
Specifies the quantization mode for model optimization and export. |
QuantizationStrategy |
Specifies the quantization strategy (algorithm) to use for model compression. |
Qwen2Variant |
Variants for Qwen2 text encoder models. |
RLAutoMLAgentType |
Defines which reinforcement learning agent families can be explored by AutoML. |
RbfCenterInitialization |
Where RBFLayer puts its centres before training. |
RbfWidthInitialization |
How RBFLayer sets each unit’s initial width. |
ReasoningEffortLevel |
How much internal deliberation a reasoning model is asked to spend before it answers. |
ReasoningParameterAdjustment |
What a reasoning-model profile did to one request setting on its way to the provider. |
ReasoningRequestParameter |
Names one per-request chat setting so a reasoning-model profile can declare it unsupported. |
RedTeamingCategory |
A family of adversarial prompts used when red teaming the safety pipeline. |
RedactionStrategy |
Strategy for redacting detected PII from text. |
RegularizationType |
Specifies the type of regularization to apply to a machine learning model. |
RelationAggregationMethod |
Methods for aggregating multiple relation scores in Relation Networks. |
RelationModuleType |
Types of relation module architectures for Relation Networks. |
RelationalAggregation |
How an R-GCN layer normalizes the messages a node receives along one relation type: the c_{i,r} of Schlichtkrull et al. |
ResNetVariant |
Defines the available ResNet (Residual Network) architecture variants. |
ReverseDiffusionVariance |
Selects the fixed variance sigma_t^2 used by a Gaussian reverse-diffusion step. |
SAGEAggregatorType |
Aggregation function type for GraphSAGE. |
SAM21ModelSize |
Defines the backbone size variants for SAM 2.1 (Segment Anything Model 2.1). |
SAMHQModelSize |
Defines the backbone size variants for SAM-HQ (High-Quality Segment Anything Model). |
SAMModelSize |
Defines the backbone size variants for SAM (Segment Anything Model). |
SANAVariant |
Variants for the SANA model family. |
SD3Variant |
Variants for the Stable Diffusion 3 model family. |
SEATSAlgorithmType |
Represents different algorithm types for SEATS (Seasonal Extraction in ARIMA Time Series) decomposition. |
SEEMModelSize |
Defines the backbone size variants for SEEM (Segment Everything Everywhere All at Once). |
SSAAlgorithmType |
Represents different algorithm types for Singular Spectrum Analysis (SSA). |
STLAlgorithmType |
Represents different algorithm types for Seasonal-Trend decomposition using LOESS (STL). |
SacActorHead |
Selects how a Soft Actor-Critic policy produces the SPREAD of its action distribution — the standard deviation of the Gaussian it samples from. |
SacCriticAggregate |
How a SAC policy update combines its two soft Q-functions. |
SacTemperatureParameter |
What a SAC agent with a learned temperature takes gradient steps on. |
SacVariant |
Which of the two Soft Actor-Critic papers a SAC agent follows. |
SafetyAction |
Defines the action to take when a safety violation is detected. |
SafetyCategory |
Comprehensive taxonomy of safety and harm categories for content classification. |
SafetySeverity |
Indicates the severity level of a safety finding. |
SamplingType |
Specifies the method used to sample or combine values when reducing data dimensions. |
SchurAlgorithmType |
Represents different algorithm types for computing the Schur decomposition of matrices. |
SeACoTrainingStage |
Which parameter group a SeACo-Paraformer training step updates. |
SegFormerModelSize |
Defines the model size variants for SegFormer (Mix Transformer backbone). |
SegGPTModelSize |
Defines the backbone size variants for SegGPT (interactive/in-context segmentation). |
SegNeXtModelSize |
Defines the model size variants for SegNeXt (Multi-Scale Convolutional Attention backbone). |
SelectionMethod |
Methods for selecting individuals for reproduction. |
SelfSupervisedLearningMethodCategory |
Categorizes self-supervised learning methods by their learning paradigm. |
SequencePoolingMode |
Strategy for collapsing a transformer encoder’s [batch, seq, dim] hidden states into a single [batch, dim] vector before the classification head, when the task is single-label per sequence. |
SequentialFeatureSelectionDirection |
Defines the direction of sequential feature selection. |
SerializationFormat |
Specifies the serialization format used for the model payload within an AIMF envelope. |
SetPredictionClassificationLoss |
The classification objective used by a DETR-family set prediction loss. |
SigLIP2Variant |
Variants for SigLIP 2 text encoder models. |
SigLIPVariant |
Variants for SigLIP text encoder models. |
SimCSEType |
Defines the training paradigms for SimCSE (Simple Contrastive Learning of Sentence Embeddings). |
SpeechLlmTrainingScope |
Which parameters a training step of an LLM-based speech recognizer (a speech encoder, an adapter and a pretrained LLM) updates. |
SpikingNeuronType |
Specifies the type of spiking neuron model to use in neuromorphic computing simulations. |
SplitCriterion |
Specifies the criterion used to determine the best way to split data in decision trees and other tree-based models. |
StepwiseMethod |
Specifies the direction of feature selection in stepwise regression and other statistical models. |
StreamingTrainingMode |
Controls the memory-bounded streaming training path (optimizer-in-backward with 8-bit Adam state and topological-min gradient release). |
SvdAlgorithmType |
Represents different algorithm types for Singular Value Decomposition (SVD). |
SwinUNETRModelSize |
Defines the backbone size variants for Swin-UNETR medical segmentation. |
SyntheticTabularTaskType |
Defines the task type for the synthetic federated tabular benchmark suite. |
T5Variant |
Variants for the T5 text encoder used in SD3 / FLUX / Imagen pipelines. |
TabRNormalization |
The normalization layer TabR builds wherever the reference model calls Normalization(d_main): before every block but the first encoder block, before the key projection when mixer normalization is on, and at the start of the prediction hea… |
TabRNumericEmbedding |
How TabR embeds its numeric features before the encoder’s first linear map: the reference implementation’s num_embeddings. |
TakagiAlgorithmType |
Represents different algorithm types for Takagi factorization of complex symmetric matrices. |
TargetPlatform |
Target hardware platforms for model deployment and optimization. |
TeacherModelType |
Specifies the type of teacher model to use for knowledge distillation. |
TemporalAggregationType |
Specifies how frame-level features are aggregated into a single video-level representation. |
TensorAxis |
The SEMANTIC role of one axis of a tensor — what the numbers along that axis actually mean. |
TestStatisticType |
Represents different types of statistical tests used to evaluate hypotheses and determine significance in data analysis. |
TimeSeriesFoundationModelTask |
Defines the tasks that a time series foundation model can perform. |
TimeSeriesModelType |
Represents different types of time series forecasting models used for analyzing and predicting sequential data over time. |
TimeSeriesTokenizationStrategy |
Tokenization strategies for time series foundation models. |
TrainingObjectiveKind |
Identifies the mathematical objective a trainable model actually optimizes. |
TrainingStageType |
Types of training stages in a multi-stage training pipeline. |
TransUNetModelSize |
Defines the backbone size variants for TransUNet medical segmentation. |
TransformerNormPlacement |
Where a transformer block applies layer normalization relative to each residual connection. |
TransformerTaskType |
Defines the different types of tasks that transformer-based AI models can perform. |
TreeSearchStrategy |
Tree search strategies for exploring the reasoning space in Tree-of-Thoughts. |
TridiagonalAlgorithmType |
Represents different algorithm types for converting a matrix to tridiagonal form. |
UNINEXTModelSize |
Defines the backbone size variants for UNINEXT. |
UduAlgorithmType |
Represents different algorithm types for UDU’ decomposition of matrices. |
UncertaintyQuantificationMethod |
Defines the supported uncertainty quantification strategies for inference. |
UniVSModelSize |
Defines the backbone size variants for UniVS (Universal Video Segmentation). |
UnivariateScoreFunction |
Built-in univariate scoring functions used by SelectKBest, SelectPercentile and GenericUnivariateSelect (the scikit-learn name of each is given in its summary). |
UnivariateScoringFunction |
Defines the scoring functions available for univariate feature selection. |
UnsupportedChatParameterHandling |
What to do when a request carries a setting the target model does not support. |
VGGVariant |
Defines the available VGG network architecture variants. |
VMambaModelSize |
Defines the backbone size variants for VMamba visual state space model. |
VectorIndexKind |
Selects the in-memory vector-search index used to back a RAG document store. |
ViTAdapterModelSize |
Defines the model size variants for ViT-Adapter. |
ViTCoMerModelSize |
Defines the model size variants for ViT-CoMer (hybrid CNN-Transformer). |
VideoModelVariant |
Defines common model size variants for video processing models (SR, interpolation, flow, etc.). |
VideoUNetArchitectureProfile |
Topology profiles supported by the shared video U-Net predictor. |
VisionMambaModelSize |
Defines the backbone size variants for Vision Mamba (Vim) segmentation. |
WaveletAlgorithmType |
Represents different types of wavelet transform algorithms for signal processing. |
WaveletType |
Defines the different types of biorthogonal wavelets that can be used for signal processing and analysis. |
WeightFunction |
Defines different weight functions used in robust statistical methods and machine learning algorithms. |
WindowFunctionType |
Defines different window functions used in signal processing and data analysis. |
Word2VecType |
Specifies the architecture type for Word2Vec models. |
X11AlgorithmType |
Represents different variants of the X-11 seasonal adjustment algorithm used in time series analysis. |
XDecoderModelSize |
Defines the backbone size variants for X-Decoder. |
YOLO11SegModelSize |
Defines the model size variants for YOLO11-Seg instance segmentation. |
YOLO26SegModelSize |
Defines the model size variants for YOLO26-Seg instance segmentation. |
YOLOv12SegModelSize |
Defines the model size variants for YOLOv12-Seg instance segmentation. |
YOLOv8SegModelSize |
Defines the size variants for YOLOv8-Seg instance segmentation models. |
YOLOv9SegModelSize |
Defines the model size variants for YOLOv9-Seg instance segmentation. |