PowerML Python Package

Requirements

Python 3.7, 3.8, 3.9, or 3.10. Check your version using python --version or python3 --version.

Installation

pip install powerml_app

Authentication

You will need a PowerML key. To get a PowerML key, go to https://powerml.co/ and log in with your email. Contact our team if you are unable to log in and we'll add you!

Quick Start

Create a python file (e.g. powerml_test.py) with this starter code and run a prediction (e.g. python powerml_test.py)!

from powerml import PowerML
config = {"powerml": {"key": "<POWERML-KEY>"}}
powerml = PowerML(config)

# Run base model
myPrompt = "hello there" # Change me and see what I can do!
response = powerml.predict(prompt=myPrompt)
print(response)

Configuration

You configure the PowerML class by passing in a dictionary like so:

from powerml import PowerML
config = {"powerml": {"key": "<POWERML-KEY>"}}
powerml = PowerML(config)

Optional: Create a config file at ~/.powerml/configure.yaml with your PowerML key. Here's an example:

powerml:
    key: "<POWERML-KEY>"

These are default keys for the PowerML class but will be overriden by any configuration dictionary passed into the class constructor.

Usage

You can use the member functions of the PowerML class, predict and fit, to make predictions with the model and fit data to the model to improve and customize it.

You can use predict to run any prompt off the bat:

from powerml import PowerML

powerml = PowerML()

# Run base model
myPrompt = "hello there"
response = powerml.predict(prompt=myPrompt)

To fit data to the model, you can use fit as so:

# Fit model to data
myData = ["item2", "item3"]
myModel = powerml.fit(myData)

To run this fitted model, you can use predict again, specifying the new model name:

# Use new model
myModelName = myModel["model_name"]
response = powerml.predict(prompt=myPrompt, model=myModelName)

PowerML Class

The PowerML class has member functions fit and predict.

Predict

predict accepts the following arguments:

def predict(self,
            prompt: str,
            model: str = "text-davinci-003",
            stop: str = "",
            max_tokens: int = 128,
            temperature: int = 0,
            ) -> str:

predict will return a string of the model's output.

fit accepts the following arguments:

def fit(self,
        data: list[str],
        model: str = ""):

fit will return a dictionary object in the following format:

{
    "model_id":"23",
    "project_id":"None",
    "user_id":"12",
    "job_id":"89",
    "model_name":"be894276039088c5f8db3f6bfaeb19953ed9ffe55f37a847a58f9fb320d307bc",
    "job_config":"{\"type\": \"prompt_tune\", \"model_name\": \"llama\"}",
    "prompt":"item2item3{{input}}",
    "creation_time":"2022-12-20 02:19:36.519260",
    "job":{
        "job_id":"89",
        "project_id":"None",
        "user_id":"12",
        "config":"{\"type\": \"prompt_tune\", \"model_name\": \"llama\"}",
        "status":"COMPLETED",
        "name":"be894276039088c5f8db3f6bfaeb19953ed9ffe55f37a847a58f9fb320d307bc",
        "metric":"None",
        "history":"None",
        "start_time":"2022-12-20 02:19:36.369450",
        "end_time":"2022-12-20 02:19:35.837668"
    }
}

ExtractTopicsModel Class

The ExtractTopicsModel class is an example class designed to extract topics from the prompt.

Usage

To instantiate a ExtractTopicsModel.

model = ExtractTopicsModel(topics)

To customize your ExtractTopicsModel instance, you can pass it examples and topics to fit.

# Examples in json for the model to fit to, in the format:
# [
#    { "example": "Using VS here for my IDE", labels: ["vscode"] },
#    { "example": "A dashboard on Chrome", labels: ["web", "dashboard"] },
# ]
examples = get_json_examples()

# Topics, e.g. ["vscode", "web", "dashboard"]
topics = get_list_of_topics()

model.fit(examples, topics)

Now, you can run this model on new examples with predict:

new_example = "Move invite teammates page to its own base route . per designs:   This PR just moves existing views around and adds a new base route (i.e. no new functionality)"

example_topics = model.predict(new_example)

Methods

__init__ is defined as follows:

def __init__(self, config={}, model_name=None):

fit is defined as follows:

def fit(self, 
        examples: list[
            {"example": str, "labels": list[str]}
        ],
        topics: list[str],
        ):

where examples is a list of dictionaries with format {"example": str, "labels": list[str]}.

predict is defined as follows:

def predict(self, prompt: str):

CreateTopicsModel Class

The CreateTopicsModel class is an example class designed to generate topics from a list of data. This is a batch process and may take a few minutes.

Usage

docs = get_list_of_data()
learn_topics = CreateTopicsModel()
learn_topics.fit(docs, topic_type='one-word system components')
topics = learn_topics.predict()

Usage with ExtractTopicsModel

Topics can be learned by CreateTopicsModel, then used in ExtractTopicsModel.

First, get topics from CreateTopicsModel:

docs = get_list_of_data()
learn_topics = CreateTopicsModel()
learn_topics.fit(docs, topic_type='one-word system components')
topics = learn_topics.predict()

Then, use ExtractTopicsModel as you normally would (as above) to fit it to examples, and then predict on new examples:

topic_model = ExtractTopicsModel()
labeled_data = get_formatted_examples()
topic_model.fit(labeled_data, topics)

new_example = "Move invite teammates page to its own base route . per designs:   This PR just moves existing views around and adds a new base route (i.e. no new functionality)"
new_example_topics = topic_model.predict(new_example)

Methods

__init__ is defined as follows:

def __init__(
        self,
        config={},
        max_output_tokens=256,):

fit is defined as follows:

def fit(self, documents: Dict[str,str], topic_type: str):

where documents is a list of strings.

predict is defined as follows:

def predict(self):

and returns a list of dictionaries with format {"name": str, “score”: float, "keywords": list[str]}.

SummarizeTopicModel Class

The SummarizeTopicModel class is an example class designed to generate topics from a list of data. This is a batch process and may take a few minutes.

Usage

model = SummarizeTopicModel()
summary = model.predict(topic_name, documents)

Methods

__init__ is defined as follows:

def __init__(
        self,
        config={},):

predict is defined as follows:

def predict(self, topic: str, documents: List[str],) -> str:

and returns a string summary of the topic, using the documents provided.

WriteEmailModel Class

The WriteEmailModel class is an example class designed to generate emails from a subject.

Usage

model = WriteEmailModel()
email = model.predict("Toys'r'Us")

ExtractMenuItemsModel Class

The ExtractMenuItemsModel class is an example class designed to generate orders from a conversation.

Usage

model = ExtractMenuItemsModel()
items = model.predict("1 Chicken Burrito")

ForecastSequenceModel Class

The ForecastSequenceModel class is an example class designed to generate a numeric sequence from a title.

Usage

model = ForecastSequenceModel()
autocompletion = model.predict("Freakonomics Radio")

AutocompleteSQLModel Class

The AutocompleteSQLModel class is an example class designed to generate sql completions from a prompt.

Usage

model = AutocompleteSQLModel()
model.fit(
    table_schemas=[
        "CREATE TABLE users ( id SERIAL PRIMARY KEY, first_name TEXT, last_name TEXT);"
    ],
    example_queries=[
        "SELECT * FROM users WHERE id=?"
    ])
autocompletion = model.predict("select * from ")

QuestionAnswerModel Class

The QuestionAnswerModel class is an example class designed to generate questions from study material.

Usage

model = QuestionAnswerModel()
note = "Greek Philosophy"
examples = [
    {'note': note,
    'questions': [{'Q': "Who said 'The only true wisdom is in knowing you know nothing.'", 'A': 'Socrates'},
    {'Q': "Who said 'I am the wisest man alive, for I know one thing, and that is that I know nothing.'", 'A': 'Socrates'}]}
]
model.fit(examples)
num_questions = 1
question_and_answer = model.predict(note, num_questions)