Class LoraFinetuning

Namespace: LMKit.Finetuning

Assembly: LM-Kit.NET.dll

Provides an engine specifically designed to fine-tune existing models using the Low-Rank Adaptation (LoRA) technique.

public sealed class LoraFinetuning : IDisposable

Inheritance: object

LoraFinetuning

Implements: IDisposable

Inherited Members: object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

The fine-tuning process involves training an existing model using a training dataset, enabling the adaptation of pre-trained models to specific tasks or data.

Constructors

LoraFinetuning(LLM, FinetuningIntent): Initializes a new instance of the LoraFinetuning class using a specified model.

LoraFinetuning(string, FinetuningIntent): Initializes a new instance of the LoraFinetuning class using specified paths.

Properties

BatchSize: Gets or sets the size of the batches used for parallel training.

ContextSize: Gets or sets the context size used during model training.

EnableSampleProcessing: Gets or sets a value indicating whether imported samples should be pre-processed by the engine. When enabled, preprocessing optimizes prompt formatting.

Iterations: Gets or sets the number of iterations that the Adam optimization algorithm performs on each training batch.

LoraTrainingParameters: Provides a reference to the parameters used for training and fine-tuning AI models with the Low-Rank Adaptation (LoRA) approach.

SampleAvgLength: Gets the average length, in tokens, of the longest sample within the training data.

SampleCount: Gets the number of training samples derived from the training data.

SampleMaxLength: Gets the length, in tokens, of the longest sample within the training data.

SampleMinLength: Gets the length, in tokens, of the shortest sample within the training data.

ThreadCount: Gets or sets the number of threads to be used for processing.
Ensures that the thread count is always at least 1 to prevent invalid configurations.

TrainingCheckpoint: Gets or sets the file path of the training checkpoint used to resume a LoRA training session.

TrainingSeed: Gets or sets the seed value used for training.
This seed is utilized to randomize the order of samples during the training process, ensuring varied training sequences.

UseGradientCheckpointing: Determines whether gradient checkpointing is enabled. Gradient checkpointing can reduce memory usage by approximately 50% at the cost of increased runtime. Disabling checkpointing may accelerate fine-tuning if sufficient RAM is available.

Methods

CheckpointToLora(LLM, string, string): Converts a LoRA training checkpoint to a LoRA format and saves it to the specified path.

ClearTrainingData(): Clears the training data.

Dispose(): Ensures the release of this instance and the complete removal of all associated unmanaged resources.

FilterSamplesBySize(int, int): Filters the training samples by size, removing those that do not fall within the specified range.

Finetune2Lora(string): Initiates the fine-tuning process using specified training parameters and model weights.

Finetune2Model(string, float, MetadataCollection): Executes the fine-tuning process and merges the resulting LoRA adapter into a new model.

GetSample(int): Retrieves a training sample from the dataset based on the provided sample index.

LoadTrainingDataFromChatHistory(ChatHistory): Loads a training dataset from a ChatHistory object. Extracts training data from a provided ChatHistory object and prepares it for the training process.

LoadTrainingDataFromText(Stream, string, Encoding): Loads a training dataset from a plain text file. This method reads a text file containing training data samples, processes the data, and prepares it for the training process.

LoadTrainingDataFromText(string, string, Encoding): Loads a training dataset from a plain text file.
Reads a text file containing training data samples separated by a specified delimiter, processes the data, and prepares it for the training process.

RemoveSample(int): Removes a training sample from the dataset based on the provided sample index.

SaveTrainingData(string, bool, Encoding, string): Saves the training data to the specified file path.

Events