Parameter-Efficient Fine-Tuning

What is Parameter-Efficient Fine-Tuning?

A set of methods that update only a small subset of a model's parameters during training, significantly reducing memory and compute requirements.

In other languages

한국어매개변수 효율적 파인튜닝

모델의 모든 매개변수를 수정하는 대신 일부 매개변수만 조정하여 적은 자원으로 효율적으로 모델을 최적화하는 기법.

日本語パラメータ効率的なファインチューニング

モデルの全パラメータを更新するのではなく、一部のパラメータのみを効率的に調整することで、少ないリソースで追加学習を行う手法。

Related Terms

• PEFTTechnique for adapting large pre-trained models by updating only a small subset of parameters, reducing memory and compute requirements
• Low-Rank AdaptationA technique for fine-tuning large models by only updating a small subset of parameters, significantly reducing computational and memory requirements.
• Full Parameter Fine-TuningTraining approach that updates all model weights for a downstream task rather than adding separate adapter layers
• Parametric TuningProcess of adjusting the internal numerical values or parameters of a model to optimize its output performance
• Soft Prompt TuningA method of adapting a model to specific tasks by learning small, adjustable vectors (prompts) without changing the core weights of the model.
• Parameter-efficient adaptersSmall, trainable modules added to a frozen pre-trained model to adapt it for specific tasks without modifying original weights
• PruningThe process of removing unnecessary or redundant parameters from a model to decrease its size and increase processing speed without significantly sacrificing functionality.
• Few-shotA machine learning setting where a model is given a small number of examples to adapt to a new task during inference without further parameter updates
• Hyperparameter OptimizationProcess of finding the best configuration of parameters that govern the training behavior of a machine learning model
• Full-rank SFTA comprehensive fine-tuning method that modifies all internal parameters of a neural network to maximize learning on new datasets.
• LoRALow-Rank Adaptation - a parameter-efficient fine-tuning method that adapts large models by only training a small number of additional weights.
• Data efficiencyThe ability of a model to achieve high performance while requiring minimal amounts of training data