opendatagen 0.0.5

Steerable data generation system for model training

⬜️ Open Datagen ⬜️

Open Datagen, a steerable data generation system for ML models training.

Features

• Generate data in the format you want
• Create custom templates with Pydantic models
• Use predefined templates

Installation

pip install --upgrade opendatagen

Setting up the OpenAI API key

export OPENAI_API_KEY='your_openai_api_key'

Usage

Example: If you want to train a small model to write great python code

variation_model = OpenAIModel(model_name=ModelName.GPT_35_TURBO_CHAT)
completion_model = OpenAIModel(model_name=ModelName.GPT_35_TURBO_INSTRUCT)

generator = DataGenerator(variation_model, completion_model)

# Create the custom template using the Pydantic models
user_template = Template(
    description="Custom template for Python exercises",
    prompt="Pthon exercice statement: {python_exercice_statement}",
    completion="Answer:\n{python_code}",
    prompt_variation_number=1,
    prompt_variables={
        "python_exercice_statement": Variable(
            name="Python exercice statement",
            temperature=1,
            max_tokens=120,
            generation_number=10
        )
    },
    completion_variables={
        "python_code": Variable(
            name="Python code",
            temperature=0,
            max_tokens=256,
            generation_number=1
        )
    }
)

data = generator.generate_data(template=user_template, 
                        output_path="output.csv")

print(data)

This code will generate a dataset of 5 medium-level Python exercises/answers formatted as you asked for.

Predefined Templates:

variation_model = OpenAIModel(model_name=ModelName.GPT_35_TURBO_CHAT)
completion_model = OpenAIModel(model_name=ModelName.GPT_35_TURBO_INSTRUCT)

generator = DataGenerator(variation_model, completion_model)

manager = TemplateManager()
template = manager.get_template(TemplateName.PRODUCT_REVIEW)

data = generator.generate_data(template=template, output_path="output.csv")

print(data)

You can find the templates in the template.json file.

Roadmap

• Enhance completion quality with sources like Internet, local files, and vector databases
• Augment and replicate sourced data
• Ensure data anonymity & open-source model support
• Future releases to support multimodal data

Note

Please note that opendatagen is initially powered by OpenAI's models. Be aware of potential biases and use the start_with and note field to guide outputs.

Acknowledgements

We would like to express our gratitude to the following open source projects and individuals that have inspired and helped us:

• Textbook Generation by VikParuchuri

• Evol-Instruct Paper (Read the paper) by WizardLM_AI

Connect

Reach us on Twitter: @thoddnn.