Fine-tune a foundation model

Foundation models are computationally expensive and trained on a large, unlabeled corpus. Fine-tuning a pre-trained foundation model is an affordable way to take advantage of their broad capabilities while customizing a model on your own small, corpus. Fine-tuning is a customization method that involved further training and does change the weights of your model.

Fine-tuning might be useful to you if you need:

• to customize your model to specific business needs

• your model to successfully work with domain-specific language, such as industry jargon, technical terms, or other specialized vocabulary

• enhanced performance for specific tasks

• accurate, relative, and context-aware responses in applications

• responses that are more factual, less toxic, and better-aligned to specific requirements

There are two main approaches that you can take for fine-tuning depending on your use case and chosen foundation model.

1. If you're interested in fine-tuning your model on domain-specific data, see Domain adaptation fine-tuning.

2. If you're interested in instruction-based fine-tuning using prompt and response examples, see Instruction-based fine-tuning.

Foundation models available for fine-tuning

You can fine-tune any of the following JumpStart foundation models:

• Bloom 3B

• Bloom 7B1

• BloomZ 3B FP16

• BloomZ 7B1 FP16

• Code Llama 13B

• Code Llama 13B Python

• Code Llama 34B

• Code Llama 34B Python

• Code Llama 70B

• Code Llama 70B Python

• Code Llama 7B

• Code Llama 7B Python

• CyberAgentLM2-7B-Chat (CALM2-7B-Chat)

• Falcon 40B BF16

• Falcon 40B Instruct BF16

• Falcon 7B BF16

• Falcon 7B Instruct BF16

• Flan-T5 Base

• Flan-T5 Large

• Flan-T5 Small

• Flan-T5 XL

• Flan-T5 XXL

• Gemma 2B

• Gemma 2B Instruct

• Gemma 7B

• Gemma 7B Instruct

• GPT-2 XL

• GPT-J 6B

• GPT-Neo 1.3B

• GPT-Neo 125M

• GPT-NEO 2.7B

• LightGPT Instruct 6B

• Llama 2 13B

• Llama 2 13B Chat

• Llama 2 13B Neuron

• Llama 2 70B

• Llama 2 70B Chat

• Llama 2 7B

• Llama 2 7B Chat

• Llama 2 7B Neuron

• Mistral 7B

• Mixtral 8x7B

• Mixtral 8x7B Instruct

• RedPajama INCITE Base 3B V1

• RedPajama INCITE Base 7B V1

• RedPajama INCITE Chat 3B V1

• RedPajama INCITE Chat 7B V1

• RedPajama INCITE Instruct 3B V1

• RedPajama INCITE Instruct 7B V1

• Stable Diffusion 2.1

Commonly supported fine-tuning hyperparameters

Different foundation models support different hyperparameters when fine-tuning. The following are commonly-supported hyperparameters that can further customize your model during training:

Inference Parameter	Description
epoch	The number of passes that the model takes through the fine-tuning dataset during training. Must be an integer greater than 1.
learning_rate	The rate at which the model weights are updated after working through each batch of fine-tuning training examples. Must be a positive float greater than 0.
instruction_tuned	Whether to instruction-train the model or not. Must be 'True' or 'False'.
per_device_train_batch_size	The batch size per GPU core or CPU for training. Must be a positive integer.
per_device_eval_batch_size	The batch size per GPU core or CPU for evaluation. Must be a positive integer.
max_train_samples	For debugging purposes or quicker training, truncate the number of training examples to this value. Value -1 means that the model uses all of the training samples. Must be a positive integer or -1.
max_val_samples	For debugging purposes or quicker training, truncate the number of validation examples to this value. Value -1 means that the model uses all of the validation samples. Must be a positive integer or -1.
max_input_length	Maximum total input sequence length after tokenization. Sequences longer than this will be truncated. If -1, max_input_length is set to the minimum of 1024 and the model_max_length defined by the tokenizer. If set to a positive value, max_input_length is set to the minimum of the provided value and the model_max_length defined by the tokenizer. Must be a positive integer or -1.
validation_split_ratio	If there is no validation channel, ratio of train-validation split from the training data. Must be between 0 and 1.
train_data_split_seed	If validation data is not present, this fixes the random splitting of the input training data to training and validation data used by the model. Must be an integer.
preprocessing_num_workers	The number of processes to use for the pre-processing. If None, main process is used for pre-processing.
lora_r	Low-rank adaptation (LoRA) r value, which acts as the scaling factor for weight updates. Must be a positive integer.
lora_alpha	Low-rank adaptation (LoRA) alpha value, which acts as the scaling factor for weight updates. Generally 2 to 4 times the size of lora_r. Must be a positive integer.
lora_dropout	Dropout value for low-rank adaptation (LoRA) layers Must be a positive float between 0 and 1.
int8_quantization	If True, model is loaded with 8 bit precision for training.
enable_fsdp	If True, training uses Fully Sharded Data Parallelism.

You can specify hyperparameter values when you fine-tune your model in Studio. For more information, see Fine-tune foundation models in Studio.

You can also override default hyperparameter values when fine-tuning your model using the SageMaker Python SDK. For more information, see Fine-tune publicly available foundation models with the JumpStartEstimator class.