Customization Config

Refer to the following table to understand the fields and settings of a given customization configuration available to your organization.

Customization configurations are created by your organization admins and define the initial parameters for a customization job, such as:

• The base model to customize

• The training methods supported

• The fine-tuning approaches supported

• The precision format

• The number of GPUs and compute nodes to use

• The parallelization techniques employed for the distributed training

• The size of micro-batches for training

• The maximum sequence length for input

An organization can have multiple customization configurations, each with different settings.

Note

For parameters that you can set at the customization job level, see the Hyperparameters reference.

---

Schema

Field	Description
name	Unique identifier for the config
base_model	The base model to customize
training_types	List of supported training methods
finetuning_types	List of supported fine-tuning approaches
precision	Model precision format (bfloat16)
num_gpus	Number of GPUs to use for training
num_nodes	Number of compute nodes to use
micro_batch_size	Size of micro-batches for training
tensor_parallel_size	Degree of tensor parallelism
use_sequence_parallel	Enables sequence parallelism
max_seq_length	Maximum sequence length for input
custom_fields	Optional custom configuration parameters
training_options	Resource configuration for the training option