TrainingPipelineConfiguration<T, TInput, TOutput>

Options & Configuration · AiDotNet.Models.Options

Configuration for a multi-step training pipeline with customizable stages.

For Beginners

Think of this as a recipe with multiple cooking steps. Just like you might marinate, then sear, then bake - training can have multiple phases where each phase teaches the model something different.

How It Works

A training pipeline defines a sequence of training stages that are executed in order. Each stage can have its own training method, optimizer, learning rate, dataset, and evaluation criteria. This enables advanced training workflows like:

Constructors

Constructor Summary
TrainingPipelineConfiguration()

Properties

Property Summary
Action<Exception, TrainingStageResult<T, TInput, TOutput>> OnPipelineError { get; set; } Gets or sets callback actions to execute on pipeline failure.
Action<List<TrainingStageResult<T, TInput, TOutput>>> OnPipelineComplete { get; set; } Gets or sets callback actions to execute when the pipeline completes.
Action<TrainingPipelineConfiguration<T, TInput, TOutput>> OnPipelineStart { get; set; } Gets or sets callback actions to execute before the pipeline starts.
Dictionary<string, object> Metadata { get; set; } Gets or sets custom metadata for the pipeline.
FineTuningData<T, TInput, TOutput> GlobalEvaluationData { get; set; } Gets or sets the evaluation data to use across all stages.
GlobalEarlyStoppingConfig GlobalEarlyStopping { get; set; } Gets or sets the global early stopping configuration applied across stages.
List<Action<TrainingStageResult<T, TInput, TOutput>>> InterStageCallbacks { get; set; } Gets or sets callback actions to execute between stages.
List<TrainingStage<T, TInput, TOutput>> Stages { get; set; } Gets or sets the ordered list of training stages in the pipeline.
MixedPrecisionType MixedPrecisionDType { get; set; } Gets or sets the mixed precision data type.
OptimizerType DefaultOptimizer { get; set; } Gets or sets the default optimizer type for all stages.
bool EnableAutoSelection { get; set; } Gets or sets whether to use automatic pipeline selection when no stages are defined.
bool EnableExperimentTracking { get; set; } Gets or sets whether to log to WandB or similar experiment trackers.
bool EvaluateAfterEachStage { get; set; } Gets or sets whether to run evaluation after each stage.
bool ResumeFromCheckpoint { get; set; } Gets or sets whether to resume from the latest checkpoint.
bool SaveIntermediateCheckpoints { get; set; } Gets or sets whether to save checkpoints between stages.
bool UseMixedPrecision { get; set; } Gets or sets whether to use mixed precision training globally.
bool VerboseLogging { get; set; } Gets or sets whether to enable verbose logging.
double DefaultLearningRate { get; set; } Gets or sets the default learning rate for all stages.
int DefaultBatchSize { get; set; } Gets or sets the default batch size for all stages.
int GlobalSeed { get; set; } Gets or sets the global random seed for reproducibility across all stages.
int MaxCheckpointsToKeep { get; set; } Gets or sets the maximum number of checkpoints to keep.
string CheckpointDirectory { get; set; } Gets or sets the directory for intermediate checkpoints.
string Description { get; set; } Gets or sets the description of this pipeline.
string ExperimentName { get; set; } Gets or sets the experiment name for tracking.
string LogDirectory { get; set; } Gets or sets the logging directory.
string Name { get; set; } Gets or sets the name of this pipeline for identification.
string ResumeCheckpointPath { get; set; } Gets or sets the specific checkpoint path to resume from.
string[] EvaluationMetrics { get; set; } Gets or sets the evaluation metrics to track.
string[] Tags { get; set; } Gets or sets tags for categorizing the pipeline.

Methods

Method Summary
List<string> Validate() Validates the pipeline configuration.
TrainingPipelineConfiguration<T, TInput, TOutput> AddAdapterMergingStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an adapter merging stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddAgenticStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an agentic behavior training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddCPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Contrastive Preference Optimization (CPO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddChainOfThoughtStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a chain-of-thought training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddCheckpointStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a checkpoint stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddCodeFineTuningStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a code fine-tuning stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddConstitutionalAIStage(string[] principles, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Constitutional AI stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddCustomStage(string name, Func<IFullModel<T, TInput, TOutput>, FineTuningData<T, TInput, TOutput>, CancellationToken, Task<IFullModel<T, TInput, TOutput>>> trainFunc, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a custom training stage with user-defined logic.
TrainingPipelineConfiguration<T, TInput, TOutput> AddDPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Direct Preference Optimization (DPO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddDistillationStage(IFullModel<T, TInput, TOutput> teacherModel, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a knowledge distillation stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddEvaluationStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an evaluation-only stage (no training, just metrics).
TrainingPipelineConfiguration<T, TInput, TOutput> AddGRPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a GRPO (Group Relative Policy Optimization) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddHarmlessnessStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a harmlessness training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddHelpfulnessStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a helpfulness training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddIPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an Identity Preference Optimization (IPO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddInstructionTuningStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an instruction tuning stage (specialized SFT).
TrainingPipelineConfiguration<T, TInput, TOutput> AddKTOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Kahneman-Tversky Optimization (KTO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddLoRAStage(int rank, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a LoRA adapter training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddMathReasoningStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a math reasoning fine-tuning stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddMultiTurnConversationStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a multi-turn conversation training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddORPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an Odds Ratio Preference Optimization (ORPO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddPPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a PPO stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddPROStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Preference Ranking Optimization (PRO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddPreferenceStage(FineTuningMethodType method, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a generic preference optimization stage with configurable method.
TrainingPipelineConfiguration<T, TInput, TOutput> AddProcessRewardModelStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a process reward model (PRM) training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddQLoRAStage(int rank, int bits, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a QLoRA (quantized LoRA) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddRLAIFStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an RLAIF (RL from AI Feedback) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddRLHFStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds an RLHF (PPO-based) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddRLStage(FineTuningMethodType method, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a generic reinforcement learning stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddRRHFStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Rank Responses to align Human Feedback (RRHF) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddRejectionSamplingStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Rejection Sampling Optimization (RSO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddRewardModelStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a reward model training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddRobustDPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Robust DPO (R-DPO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddSFTStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a supervised fine-tuning (SFT) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddSLiCStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Sequence Likelihood Calibration (SLiC-HF) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddSPINStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Self-Play Fine-Tuning (SPIN) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddSafetyAlignmentStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a safety alignment stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddSelfRewardingStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a self-rewarding stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddSimPOStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a Simple Preference Optimization (SimPO) stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddStage(TrainingStage<T, TInput, TOutput> stage) Adds a training stage to the pipeline.
TrainingPipelineConfiguration<T, TInput, TOutput> AddSyntheticDataStage(IFullModel<T, TInput, TOutput> teacherModel, Action<TrainingStage<T, TInput, TOutput>> configure) Adds a synthetic data training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AddToolUseStage(Action<TrainingStage<T, TInput, TOutput>> configure) Adds a tool use training stage.
TrainingPipelineConfiguration<T, TInput, TOutput> AgentTraining(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> toolData, FineTuningData<T, TInput, TOutput> agenticData) Creates an agent/tool-use training pipeline.
TrainingPipelineConfiguration<T, TInput, TOutput> AnthropicClaude(FineTuningData<T, TInput, TOutput> sftData, string[] principles) Creates an Anthropic Claude-style pipeline (SFT → Constitutional AI → RLHF).
TrainingPipelineConfiguration<T, TInput, TOutput> Auto(FineTuningData<T, TInput, TOutput> availableData) Automatically selects an appropriate pipeline based on available data.
TrainingPipelineConfiguration<T, TInput, TOutput> CodeModel(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> codeExecutionData) Creates a code model training pipeline.
TrainingPipelineConfiguration<T, TInput, TOutput> ConstitutionalAI(FineTuningData<T, TInput, TOutput> sftData, string[] principles) Creates a Constitutional AI pipeline (SFT → CAI critique/revision → preference).
TrainingPipelineConfiguration<T, TInput, TOutput> CurriculumLearning(ValueTuple<string, FineTuningData<T, TInput, TOutput>>[] curriculumStages) Creates a curriculum learning pipeline with progressively harder stages.
TrainingPipelineConfiguration<T, TInput, TOutput> DeepSeek(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> grpoData) Creates a DeepSeek-style pipeline (SFT → GRPO).
TrainingPipelineConfiguration<T, TInput, TOutput> FullRLHF(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> rlData) Creates a full RLHF pipeline (SFT → Reward Model → PPO).
TrainingPipelineConfiguration<T, TInput, TOutput> InstructGPT(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> preferenceData, FineTuningData<T, TInput, TOutput> rlData) Creates an OpenAI InstructGPT-style pipeline (SFT → Reward Model → PPO).
TrainingPipelineConfiguration<T, TInput, TOutput> IterativeRefinement(int iterations, FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> preferenceData) Creates an iterative refinement pipeline that runs multiple DPO rounds.
TrainingPipelineConfiguration<T, TInput, TOutput> IterativeSPIN(int iterations, FineTuningData<T, TInput, TOutput> sftData) Creates an iterative SPIN pipeline.
TrainingPipelineConfiguration<T, TInput, TOutput> KnowledgeDistillation(IFullModel<T, TInput, TOutput> teacherModel, FineTuningData<T, TInput, TOutput> data) Creates a distillation pipeline (teacher → student).
TrainingPipelineConfiguration<T, TInput, TOutput> Llama2(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> preferenceData) Creates a Meta Llama 2-style pipeline (SFT → Rejection Sampling → DPO).
TrainingPipelineConfiguration<T, TInput, TOutput> LoRAFineTuning(int rank, FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> preferenceData) Creates a memory-efficient LoRA fine-tuning pipeline.
TrainingPipelineConfiguration<T, TInput, TOutput> MathReasoning(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> cotData) Creates a math reasoning model pipeline.
TrainingPipelineConfiguration<T, TInput, TOutput> ORPOAlignment(FineTuningData<T, TInput, TOutput> data) Creates a reference-free alignment pipeline using ORPO.
TrainingPipelineConfiguration<T, TInput, TOutput> QLoRAFineTuning(int rank, int bits, FineTuningData<T, TInput, TOutput> data) Creates a QLoRA fine-tuning pipeline for maximum memory efficiency.
TrainingPipelineConfiguration<T, TInput, TOutput> SafetyFocused(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> safetyData, string[] principles) Creates a safety-focused training pipeline.
TrainingPipelineConfiguration<T, TInput, TOutput> SimPOAlignment(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> preferenceData) Creates a SimPO alignment pipeline (reference-free, simple).
TrainingPipelineConfiguration<T, TInput, TOutput> StandardAlignment(FineTuningData<T, TInput, TOutput> sftData, FineTuningData<T, TInput, TOutput> preferenceData) Creates a standard SFT → DPO pipeline (most common alignment workflow).
void ValidateOrThrow() Throws an exception if the pipeline is invalid.