Finetune Any LLM with SFT

Supervised Finetuning with LLMs¶

Large Language Models or LLMs have seen a lot of progress in the last year. We went from now ChatGPT competitor to a whole zoo of LLMs, including Meta AI's Llama 3, Mistrals Mistral & Mixtral models, TII Falcon, and many more. Those LLMs can be used for a variety of tasks, including chatbots, question answering, summarization without any additional training. However, if you want to customize a model for your application. You may need to fine-tune the model on your data to achieve higher quality results than prompting or saving cost by training smaller models more efficient model.

This blog post walks you thorugh how to fine-tune open LLMs using Hugging Face TRL, Transformers & datasets in 2024. In the blog, we are going to:

1. Define our use case
2. Setup development environment
3. Create and prepare the dataset
4. Fine-tune LLM using trl and the SFTTrainer
5. Test and evaluate the LLM
6. Deploy the LLM for Production

Note: This blog was created to run on consumer size GPUs (24GB), e.g. NVIDIA A10G or RTX 4090/3090, but can be easily adapted to run on bigger GPUs.

1. Define our use case¶

When fine-tuning LLMs, it is important you know your use case and the task you want to solve. This will help you to choose the right model or help you to create a dataset to fine-tune your model. If you haven't defined your use case yet. You might want to go back to the drawing board. I want to mention that not all use cases require fine-tuning and it is always recommended to evaluate and try out already fine-tuned models or API-based models before fine-tuning your own model.

As an example, we are going to use the following use case:

We want to fine-tune a model, which can generate SQL queries based on a natural language instruction, which can then be integrated into our BI tool. The goal is to reduce the time it takes to create a SQL query and make it easier for non-technical users to create SQL queries.

Text to SQL can be a good use case for fine-tuning LLMs, as it is a complex task that requires a lot of (internal) knowledge about the data and the SQL language.

2. Setup development environment¶

Our first step is to install Hugging Face Libraries and Pyroch, including trl, transformers and datasets. If you haven't heard of trl yet, don't worry. It is a new library on top of transformers and datasets, which makes it easier to fine-tune, rlhf, align open LLMs.

If you are using a GPU with Ampere architecture (e.g. NVIDIA A10G or RTX 4090/3090) or newer you can use Flash attention. Flash Attention is a an method that reorders the attention computation and leverages classical techniques (tiling, recomputation) to significantly speed it up and reduce memory usage from quadratic to linear in sequence length. The TL;DR; accelerates training up to 3x. Learn more at FlashAttention.

Note: If your machine has less than 96GB of RAM and lots of CPU cores, reduce the number of MAX_JOBS. On the g6.2xlarge we used 4.

Installing flash attention can take quite a bit of time (10-45 minutes).

We will use the Hugging Face Hub as a remote model versioning service. This means we will automatically push our model, logs and information to the Hub during training. You must register on the Hugging Face for this. After you have an account, we will use the login util from the huggingface_hub package to log into our account and store our token (access key) on the disk.

3. Create and prepare the dataset¶

Once you have determined that fine-tuning is the right solution we need to create a dataset to fine-tune our model. The dataset should be a diverse set of demonstrations of the task you want to solve. There are several ways to create such a dataset, including:

• Using existing open-source datasets, e.g., Spider
• Using LLMs to create synthetically datasets, e.g., Alpaca
• Using Humans to create datasets, e.g., Dolly.
• Using a combination of the above methods, e.g., Orca

Each of the methods has its own advantages and disadvantages and depends on the budget, time, and quality requirements. For example, using an existing dataset is the easiest but might not be tailored to your specific use case, while using humans might be the most accurate but can be time-consuming and expensive. It is also possible to combine several methods to create an instruction dataset, as shown in Orca: Progressive Learning from Complex Explanation Traces of GPT-4.

In our example we will use an already existing dataset called sql-create-context, which contains samples of natural language instructions, schema definitions and the corresponding SQL query.

With the latest release of trl we now support popular instruction and conversation dataset formats. This means we only need to convert our dataset to one of the supported formats and trl will take care of the rest. Those formats include:

• conversational format

{"messages": [{"role": "system", "content": "You are..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
{"messages": [{"role": "system", "content": "You are..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
{"messages": [{"role": "system", "content": "You are..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

• instruction format

{"prompt": "<prompt text>", "completion": "<ideal generated text>"}
{"prompt": "<prompt text>", "completion": "<ideal generated text>"}
{"prompt": "<prompt text>", "completion": "<ideal generated text>"}

In our example we are going to load our open-source dataset using the 🤗 Datasets library and then convert it into the the conversational format, where we include the schema definition in the system message for our assistant. We'll then save the dataset as jsonl file, which we can then use to fine-tune our model. We are randomly downsampling the dataset to only 10,000 samples.

Note: This step can be different for your use case. For example, if you have already a dataset from, e.g. working with OpenAI, you can skip this step and go directly to the fine-tuning step.

4. Fine-tune LLM using trl and the SFTTrainer¶

We are now ready to fine-tune our model. We will use the SFTTrainer from trl to fine-tune our model. The SFTTrainer makes it straightfoward to supervise fine-tune open LLMs. The SFTTrainer is a subclass of the Trainer from the transformers library and supports all the same features, including logging, evaluation, and checkpointing, but adds additiional quality of life features, including:

• Dataset formatting, including conversational and instruction format
• Training on completions only, ignoring prompts
• Packing datasets for more efficient training
• PEFT (parameter-efficient fine-tuning) support including Q-LoRA
• Preparing the model and tokenizer for conversational fine-tuning (e.g. adding special tokens)

We will use the dataset formatting, packing and PEFT features in our example. As peft method we will use QLoRA a technique to reduce the memory footprint of large language models during finetuning, without sacrificing performance by using quantization. If you want to learn more about QLoRA and how it works, check out Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA blog post.

Now, lets get started! 🚀

First, we need to load our dataset from disk.

Next, we will load our LLM. For our use case we are going to use Llama 3.1 8B. But we can easily swap out the model for another model, e.g. Mistral or Mixtral models, TII Falcon, or any other LLMs by changing our model_id variable. We will use bitsandbytes to quantize our model to 4-bit.

Note: Be aware the bigger the model the more memory it will require. In our example we will use the 8B version, which can be tuned on 24GB GPUs. If you have a smaller GPU.

Correctly, preparing the LLM and Tokenizer for training chat/conversational models is crucial. We need to add new special tokens to the tokenizer and model and teach to understand the different roles in a conversation. In trl we have a convinient method called setup_chat_format, which:

• Adds special tokens to the tokenizer, e.g. <|im_start|> and <|im_end|>, to indicate the start and end of a conversation.
• Resizes the model’s embedding layer to accommodate the new tokens.
• Sets the chat_template of the tokenizer, which is used to format the input data into a chat-like format. The default is chatml from OpenAI.

The SFTTrainer  supports a native integration with peft, which makes it super easy to efficiently tune LLMs using, e.g. QLoRA. We only need to create our LoraConfig and provide it to the trainer. Our LoraConfig parameters are defined based on the qlora paper and sebastian's blog post.

Before we can start our training we need to define the hyperparameters (TrainingArguments) we want to use.

We now have every building block we need to create our SFTTrainer to start then training our model.

Start training our model by calling the train() method on our Trainer instance. This will start the training loop and train our model for 3 epochs. Since we are using a PEFT method, we will only save the adapted model weights and not the full model.

The training with Flash Attention for 3 epochs with a dataset of 10k samples took 02:05:58 on a g6.2xlarge. The instance costs 1,212$/h which brings us to a total cost of only 1.8$.

Merge LoRA adapter in to the original model¶

When using QLoRA, we only train adapters and not the full model. This means when saving the model during training we only save the adapter weights and not the full model. If you want to save the full model, which makes it easier to use with Text Generation Inference you can merge the adapter weights into the model weights using the merge_and_unload method and then save the model with the save_pretrained method. This will save a default model, which can be used for inference.

Note: This requires > 30GB CPU Memory.

4. Test Model and run Inference¶

After the training is done we want to evaluate and test our model. We will load different samples from the original dataset and evaluate the model on those samples, using a simple loop and accuracy as our metric.

Note: Evaluating Generative AI models is not a trivial task since 1 input can have multiple correct outputs. If you want to learn more about evaluating generative models, check out Evaluate LLMs and RAG a practical example using Langchain and Hugging Face blog post.

Let’s load our test dataset try to generate an instruction.

Nice! Our model was able to generate a SQL query based on the natural language instruction. Lets evaluate our model on the full 2,500 samples of our test dataset. Note: As mentioned above, evaluating generative models is not a trivial task. In our example we used the accuracy of the generated SQL based on the ground truth SQL query as our metric. An alternative way could be to automatically execute the generated SQL query and compare the results with the ground truth. This would be a more accurate metric but requires more work to setup.

We evaluated our model on 1000 samples from the evaluation dataset and got an accuracy of 80.00%, which took ~25 minutes. This is quite good, but as mentioned you need to take this metric with a grain of salt. It would be better if we could evaluate our model by running the qureies against a real database and compare the results. Since there might be different "correct" SQL queries for the same instruction. There are also several ways on how we could improve the performance by using few-shot learning, using RAG, Self-healing to generate the SQL query.

6. Deploy the LLM for Production¶

You can now deploy your model to production. For deploying open LLMs into production we recommend using Text Generation Inference (TGI). TGI is a purpose-built solution for deploying and serving Large Language Models (LLMs). TGI enables high-performance text generation using Tensor Parallelism and continous batching for the most popular open LLMs, including Llama, Mistral, Mixtral, StarCoder, T5 and more. Text Generation Inference is used by companies as IBM, Grammarly, Uber, Deutsche Telekom, and many more. There are several ways to deploy your model, including:

• Deploy LLMs with Hugging Face Inference Endpoints
• Hugging Face LLM Inference Container for Amazon SageMaker

If you have docker installed you can use the following command to start the inference server.

Note: Make sure that you have enough GPU memory to run the container. Restart kernel to remove all allocated GPU memory from the notebook.

Once your container is running you can send requests using the openai or huggingface_hub sdk. Here we ll use the openai sdk to send a request to our inference server. If you don't have the openai sdk installed you can install it using pip install openai.

Awesome, Don't forget to stop your container once you are done.