Have you ever wanted to fine-tune a large language model, but didn't have the time or resources to train it from scratch? In this post, I'll show you how to use adapter-transformers to fine-tune Flan-T5 on the XSum dataset in five minutes. I'll also show you how to use the resulting model to generate summaries for any text.
Fine-tuning is the process of training a pre-trained model on a new dataset. This usually takes less time than training a model from scratch, but it can still take a long time.
adapter-transformers is a library for fine-tuning transformers, based on HuggingFace's transformers library. It has a number of new features that make it easier to fine-tune transformers, such as LoRA (Low Rank Adaptation), which allow you to reduce the amount of parameters that need to be trained.
LoRA works by training two small matrices that are multiplied together to produce a large matrix. This large matrix is then added to the pre-trained model's weights, usually in the attention layers. Adding the matrices is equivalent to adding the results of the matrix multiplication, which is done in training.
Flan-T5 is a series of models by Google that are based on the T5 architecture. They are text-to-text transformers that can be used for a variety of tasks, making them a great starting point for fine-tuning.
In this post, I'll be using flan-t5-base, which is one of the smaller models in the series.
XSum stands for Extreme Summarization and is a dataset of documents and summaries.
In this post, I'll be using a small subset of the XSum dataset, which contains 1,000 documents and their summaries.
I'll be using Google Colab for this post, but you can also run it locally if you have a GPU and the necessary libraries installed.
To install the libraries that aren't installed in Google Colab, you'll have to run the following commands:
In order to use the Flan-T5 model, we'll have to convert the documents and summaries into tokens that the model can understand.
To do this, we'll use the class from the library. We'll also use a prefix before the documents, which is used by the model to determine what task it should perform. In this case, the prefix is . We have to do this because this is the format the model expects.
We'll be using the library to load the dataset, and we will load a subset of the dataset from each split.
Now that we have the dataset, we can prepare the model for training.
Now that we have the model, we can prepare the trainer.
Now that we have the trainer, we can train the model. This will take a while, especially depending on your hardware.
Now that we have the trained model, we can evaluate its performance on the test split.
Like I mentioned earlier, we will need to add the matrices.
We will use the validation split of the dataset and compare our summaries with the actual summaries.
Hopefully this tutorial was helpful in showing how to quickly fine-tune a model for a new task using adapters. My source code for this tutorial can be found on Google Colab.