components prompt_crafter

Prompt Crafter

prompt_crafter

Overview

This component is used to create prompts from a given dataset. From a given jinja prompt template, it will generate prompts. It can also create few-shot prompts given a few-shot dataset and the number of shots.

Version: 0.0.12

View in Studio: https://ml.azure.com/registries/azureml/components/prompt_crafter/version/0.0.12

Inputs

Name	Description	Type	Default	Optional	Enum
prompt_type	Determine the prompt format. This component supports chat and completion models. Completions: {"prompt": "Few shot prompts will go here"} Chat: Chat models have 3 roles: System, User and Assistant. Example: {"prompt": [{"role": "system", "content": "You are a helpful assistant"}, {"role": "user", "content": "Example chat input"}, {"role": "assistant", "content": "Example chat output"}]}	string	completions	False	['chat', 'completions']
test_data	The uri file (jsonl) used to generate prompts.	uri_file		False
prompt_pattern	The pattern to be used to generate the prompts. It should be a valid jinja template. Example: Input Data : {"question":"Example Question?", "choices":{"option":["Answer1","Answer2","Answer3","Answer4"]}, "answerKey":"D"} The prompt pattern for the above input data can be: "Question: {{question}}\n Choices are: (1) {{choices.option[0]}}\n (2) {{choices.option[1]}}\n (3) {{choices.option[2]}}\n (4) {{choices.option[3]}}\n"	string		False
few_shot_pattern	The pattern used to generate the few shot portion of a prompt. It should be a valid jinja template. If this pattern is not provided, few shot prompts are generated from a concatenation of prompt_pattern and output_pattern. Example: Input Data : {"question":"Example Question?", "choices":{"option":["Answer1","Answer2","Answer3","Answer4"]}, "answerKey":"D"} The few shot pattern for the above input data can be: "Question: {{question}}\n Choices are: (1) {{choices.option[0]}}\n (2) {{choices.option[1]}}\n (3) {{choices.option[2]}}\n (4) {{choices.option[3]}}\n Answer: {{answerKey}}"	string		True
n_shots	The number of shots to use in the few-shot prompts. Default is 0, which means no few-shot examples will be generated. n_shots must be smaller than the size of few_shot dataset.	integer	0	False
few_shot_data	The uri file(jsonl) to be used to generate the n-shot prompts.	uri_file		True
output_pattern	The jinja template representing the expected output that would be used for few shot prompts when n_shot > 0. e.g: {{answerKey}}	string		False
few_shot_separator	The separator to be added between few-shot prompts.	string		True
system_message	This is the description of the task that the Assisstant should perform. Applicable for chat models only. e.g: "You are a helpful assistant." will be added to system role for chat models. {"role": "system", "content": "You are a helpful assistant"}	string		True
prefix	The prefix to be added to the prompts. e.g: "Question: "	string		True
random_seed	Random seed for sampling few-shots; if not specified, 0 is used.	integer	0	True
ground_truth_column_name	This will be used as the ground truth column if present in the input. If not present, the output_pattern will be used as the ground truth.	string		True
additional_columns	Any additional columns that would be helpful for computing metrics, if present in the input. If there're multiple such columns, they should be separated by comma (",").	string		True

Outputs

Name	Description	Type
output_file	Output file path where few_shot_prompt data will be written.	uri_file

Environment

azureml://registries/azureml/environments/model-evaluation/labels/latest