TrainingStage<T, TInput, TOutput>
Models & Types · AiDotNet.Models.Options
Represents a single stage in a training pipeline with comprehensive configuration options.
For Beginners
Think of each stage as a chapter in a training book. Each chapter teaches the model something different, and you can configure exactly how that teaching happens.
How It Works
A training stage encapsulates all configuration needed for one step in a multi-stage training pipeline. Each stage can have its own:
Constructors
| Constructor | Summary |
|---|---|
TrainingStage() |
Properties
| Property | Summary |
|---|---|
CheckpointMetricType BestCheckpointMetric { get; set; } |
Gets or sets the metric to use for determining the best checkpoint. |
Dictionary<int, int> DistillationLayerMapping { get; set; } |
Gets or sets the layer mapping for intermediate distillation. |
Dictionary<string, double> DataMixingRatios { get; set; } |
Gets or sets the data mixing ratio when combining multiple datasets. |
Dictionary<string, object> Metadata { get; set; } |
Gets or sets custom metadata for this stage. |
DistributedStrategy DistributedStrategy { get; set; } |
Gets or sets the distributed training strategy for this stage. |
EarlyStoppingConfig EarlyStopping { get; set; } |
Gets or sets early stopping configuration specific to this stage. |
FineTuningData<T, TInput, TOutput> TrainingData { get; set; } |
Gets or sets the training data for this stage. |
FineTuningData<T, TInput, TOutput> ValidationData { get; set; } |
Gets or sets the validation data for this stage. |
FineTuningMethodType FineTuningMethod { get; set; } |
Gets or sets the fine-tuning method to use in this stage. |
FineTuningOptions<T> Options { get; set; } |
Gets or sets the fine-tuning options for this stage. |
Func<FineTuningData<T, TInput, TOutput>, FineTuningData<T, TInput, TOutput>> DataPreprocessor { get; set; } |
Gets or sets custom data preprocessing for this stage. |
Func<IFullModel<T, TInput, TOutput>, FineTuningData<T, TInput, TOutput>, CancellationToken, Task<IFullModel<T, TInput, TOutput>>> CustomTrainingFunction { get; set; } |
Gets or sets the custom training function for custom stages. |
Func<IFullModel<T, TInput, TOutput>, FineTuningData<T, TInput, TOutput>, Task<Dictionary<string, double>>> CustomEvaluationFunction { get; set; } |
Gets or sets the custom evaluation function for this stage. |
Func<TrainingStageResult<T, TInput, TOutput>, bool> RunCondition { get; set; } |
Gets or sets conditions that must be met to run this stage. |
IFullModel<T, TInput, TOutput> CritiqueModel { get; set; } |
Gets or sets the model to use for generating critiques. |
IFullModel<T, TInput, TOutput> RewardModel { get; set; } |
Gets or sets the reward model to use for RLHF stages. |
IFullModel<T, TInput, TOutput> TeacherModel { get; set; } |
Gets or sets the teacher model for distillation stages. |
ILossFunction<T> CustomLossFunction { get; set; } |
Gets or sets custom loss function for this stage. |
LearningRateSchedulerType SchedulerType { get; set; } |
Gets or sets the learning rate scheduler type. |
MixedPrecisionType MixedPrecisionDType { get; set; } |
Gets or sets the mixed precision data type. |
OptimizerType OptimizerOverride { get; set; } |
Gets or sets the optimizer type override for this stage. |
PreferenceLossType PreferenceLossType { get; set; } |
Gets or sets the loss type for preference optimization. |
StageCallbacks<T, TInput, TOutput> Callbacks { get; set; } |
Gets or sets stage-specific callbacks. |
TimeSpan MaxDuration { get; set; } |
Gets or sets the maximum duration for this stage. |
TrainingStageType StageType { get; set; } |
Gets or sets the type of training stage. |
bool BestCheckpointMetricMaximize { get; set; } |
Gets or sets whether higher is better for the best checkpoint metric. |
bool Enabled { get; set; } |
Gets or sets whether this stage is enabled (skipped if false). |
bool FreezeBaseModel { get; set; } |
Gets or sets whether to freeze the base model during this stage. |
bool GRPOUseRelativeRewards { get; set; } |
Gets or sets whether to use relative rewards in GRPO. |
bool IsEvaluationOnly { get; set; } |
Gets or sets whether this stage is evaluation-only (no training). |
bool LogGradientNorms { get; set; } |
Gets or sets whether to log gradient norms. |
bool LogLearningRate { get; set; } |
Gets or sets whether to log learning rate. |
bool MergeLoRAAfterTraining { get; set; } |
Gets or sets whether to merge LoRA weights into base model after training. |
bool SaveCheckpointAfter { get; set; } |
Gets or sets whether to save a checkpoint after this stage. |
bool SaveOnlyBest { get; set; } |
Gets or sets whether to save only the best checkpoint based on validation metrics. |
bool ShareReferenceModel { get; set; } |
Gets or sets whether to share reference model with the training model. |
bool ShuffleData { get; set; } |
Gets or sets whether to shuffle the training data each epoch. |
bool SyncBatchNorm { get; set; } |
Gets or sets whether to sync batch normalization across devices. |
bool UpdateReferenceModel { get; set; } |
Gets or sets whether to update the reference model periodically. |
bool UseDeterministicAlgorithms { get; set; } |
Gets or sets whether to use deterministic algorithms (may be slower). |
bool UseDynamicLossScaling { get; set; } |
Gets or sets whether to use dynamic loss scaling for mixed precision. |
bool UseGradientCheckpointing { get; set; } |
Gets or sets whether to use gradient checkpointing to save memory. |
bool UseGradualUnfreezing { get; set; } |
Gets or sets whether to gradually unfreeze layers during training. |
bool UseIntermediateDistillation { get; set; } |
Gets or sets whether to use intermediate layer distillation. |
bool UseLoRA { get; set; } |
Gets or sets whether to use LoRA (Low-Rank Adaptation) for this stage. |
bool UseMixedPrecision { get; set; } |
Gets or sets whether to use mixed precision training (FP16/BF16). |
bool UseQLoRA { get; set; } |
Gets or sets whether to use QLoRA (quantized LoRA) for memory efficiency. |
bool UseReferenceModel { get; set; } |
Gets or sets whether to use a reference model for preference methods. |
double AdamBeta1 { get; set; } |
Gets or sets the Adam beta1 parameter (momentum). |
double AdamBeta2 { get; set; } |
Gets or sets the Adam beta2 parameter (RMSprop-like). |
double AdamEpsilon { get; set; } |
Gets or sets the Adam epsilon for numerical stability. |
double ContrastiveMargin { get; set; } |
Gets or sets the margin for contrastive preference methods. |
double DPOBeta { get; set; } |
Gets or sets the beta parameter for DPO/IPO loss. |
double DistillationAlpha { get; set; } |
Gets or sets the alpha for balancing hard vs soft targets. |
double DistillationTemperature { get; set; } |
Gets or sets the distillation temperature. |
double EntropyCoefficient { get; set; } |
Gets or sets the entropy bonus coefficient for exploration. |
double GAELambda { get; set; } |
Gets or sets the GAE lambda for advantage estimation. |
double InitialLossScale { get; set; } |
Gets or sets the initial loss scale for mixed precision. |
double KLPenaltyCoefficient { get; set; } |
Gets or sets the KL penalty coefficient for RLHF. |
double LearningRate { get; set; } |
Gets or sets the learning rate for this stage. |
double LoRAAlpha { get; set; } |
Gets or sets the LoRA alpha scaling factor. |
double LoRADropout { get; set; } |
Gets or sets the LoRA dropout rate. |
double MaxGradientNorm { get; set; } |
Gets or sets the maximum gradient norm for gradient clipping. |
double MinLearningRate { get; set; } |
Gets or sets the minimum learning rate (for schedulers with decay). |
double PPOClipRange { get; set; } |
Gets or sets the PPO clip range. |
double PreferenceLabelSmoothing { get; set; } |
Gets or sets the label smoothing factor for preference learning. |
double RejectionSamplingMinReward { get; set; } |
Gets or sets the minimum reward threshold for rejection sampling. |
double SchedulerPower { get; set; } |
Gets or sets the power for polynomial decay scheduler. |
double SelfPlayTemperature { get; set; } |
Gets or sets the generation temperature for self-play responses. |
double ValueFunctionCoefficient { get; set; } |
Gets or sets the value function coefficient for PPO. |
double WarmupRatio { get; set; } |
Gets or sets the warmup ratio (fraction of total steps for warmup). |
double WeightDecay { get; set; } |
Gets or sets the weight decay (L2 regularization) coefficient. |
int BatchSize { get; set; } |
Gets or sets the batch size for this stage. |
int CheckpointSaveEpochs { get; set; } |
Gets or sets the checkpoint save interval (in epochs). |
int CheckpointSaveSteps { get; set; } |
Gets or sets the checkpoint save interval (in steps). |
int CritiqueRevisionRounds { get; set; } |
Gets or sets the number of critique-revision rounds. |
int Epochs { get; set; } |
Gets or sets the number of epochs for this stage. |
int EvaluationSteps { get; set; } |
Gets or sets the evaluation interval (in steps). |
int GRPOGroupSize { get; set; } |
Gets or sets the group size for GRPO. |
int GradientAccumulationSteps { get; set; } |
Gets or sets the gradient accumulation steps. |
int GradualUnfreezingInterval { get; set; } |
Gets or sets the epoch interval for gradual unfreezing. |
int LoRARank { get; set; } |
Gets or sets the LoRA rank (dimension of low-rank matrices). |
int LoggingSteps { get; set; } |
Gets or sets the logging interval (in steps). |
int MaxCheckpointsToKeep { get; set; } |
Gets or sets the maximum number of checkpoints to keep. |
int MaxSteps { get; set; } |
Gets or sets the maximum number of steps for this stage. |
int NumCycles { get; set; } |
Gets or sets the number of cycles for cosine scheduler with restarts. |
int PPOEpochsPerBatch { get; set; } |
Gets or sets the number of PPO epochs per batch. |
int QLoRABits { get; set; } |
Gets or sets the quantization bits for QLoRA. |
int RandomSeed { get; set; } |
Gets or sets the random seed for this stage. |
int ReferenceModelUpdateInterval { get; set; } |
Gets or sets the interval (in steps) for updating the reference model. |
int RejectionSamplingN { get; set; } |
Gets or sets the number of samples to generate for rejection sampling. |
int RejectionSamplingTopK { get; set; } |
Gets or sets the top-K samples to keep from rejection sampling. |
int RolloutSamples { get; set; } |
Gets or sets the number of rollout samples per update. |
int SelfPlayIterations { get; set; } |
Gets or sets the number of self-play iterations. |
int SelfPlayResponsesPerPrompt { get; set; } |
Gets or sets the number of responses to generate per prompt in self-play. |
int UnfreezeTopNLayers { get; set; } |
Gets or sets the number of layers to unfreeze from the top. |
int WarmupSteps { get; set; } |
Gets or sets the number of warmup steps. |
int? DataShuffleSeed { get; set; } |
Gets or sets the random seed for data shuffling (for reproducibility). |
int? MaxTrainingSamples { get; set; } |
Gets or sets the maximum number of training samples to use. |
string CustomMetricName { get; set; } |
Gets or sets custom metric name when BestCheckpointMetric is Custom. |
string Description { get; set; } |
Gets or sets a description of what this stage accomplishes. |
string Name { get; set; } |
Gets or sets the name of this stage for logging and identification. |
string[] ConstitutionalPrinciples { get; set; } |
Gets or sets the constitutional principles for CAI stages. |
string[] FrozenLayers { get; set; } |
Gets or sets layer names/patterns to freeze during this stage. |
string[] LoRATargetModules { get; set; } |
Gets or sets which modules to apply LoRA to. |
string[] MetricsToTrack { get; set; } |
Gets or sets the metrics to track during this stage. |
string[] Tags { get; set; } |
Gets or sets tags for categorizing this stage. |
string[] TrainableLayers { get; set; } |
Gets or sets layer names/patterns to unfreeze (train) during this stage. |