components aoai_finetuning

AOAI Finetuning Job

aoai_finetuning

Overview

Upload data to Azure OpenAI resource, finetune model and delete data

Version: 0.0.3

View in Studio: https://ml.azure.com/registries/azureml/components/aoai_finetuning/version/0.0.3

Inputs

Name	Description	Type	Default	Optional	Enum
endpoint_name	The endpoint name or AOAI resource name.	string		False
endpoint_resource_group	Resource group for the AOAI resource.	string		True
endpoint_subscription	Subscription for the AOAI resource.	string		True
training_file_path	jsonl source file/folder for training dataset.	uri_file		False
validation_file_path	source file/folder for validation dataset.	uri_file		True
model	GPT model engine	string	gpt-35-turbo-0613	False
task_type	Dataset type - chat or completion	string		False	['chat', 'completion', 'embedding']
n_epochs	Number of training epochs. If not provided, it will be determined dynamically based on the input data.	integer		True
batch_size	Global batch size. If not provided, it will be determined dynamically based on the input data.	integer		True
learning_rate_multiplier	The learning rate multiplier to use for training. If not provided, it will be determined dynamically based on the input data.	number		True
suffix	A string of up to 18 characters that will be added to your fine-tuned model name	string		True
export_merged_weights	To get the merged wights of the model as output as well. Default is false	boolean		True
completion_override	To override the task type to completion. Default is false	boolean		True
full_finetune	To perform full finetuning. Default is false	boolean		True
lora_v2	To use lora V2. Default is false	boolean		True
lora_dimensions	The size of LoRA dimensions in self attention layer. If not provided, it will be determined dynamically.	integer		True
context_window	Context length of the model. If not provided, context window will be determined dynamically.	integer		True
file_spm_rate	file spm rate should be between [0,1]	number		True
weight_decay_multiplier	Weight Decay Multiplier for training. Not applicable for embedding finetuning	number		True
prompt_loss_weight	Loss weight defined on prompt (i.e. user message). Note that loss weight defined on completion (i.e. assistant message) is alwyas 1.0.	number		True
trim_mode	Trim method if data is longer than context window	string		True	['left', 'right', 'discard']
check_point_interval	Checkpointing frequency based on steps. Applicable only to embedding finetuning.	integer		True
num_steps	Total training steps. Applicable only to embedding finetuning.	integer		True
shuffle_type	Shuffle type for input train dataset. Buffer means shuffle with a small buffer.	string		True	['none', 'full', 'buffer']

Outputs

Name	Description	Type
aoai_finetuning_output	Contains finetuned model id in output file in JSON/custom class format	uri_file

Environment

azureml://registries/azureml-staging/environments/aoai-data-upload-finetune/versions/3