opendatagen 0.0.12

Steerable data generation system for LLM fine-tuning

⬜️ Open Datagen ⬜️

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

Features

• Generate synthetic datasets with a fixed format using:

  • User-defined template with pydantic object
  • Predefined templates

• Data anonymization

• Quality enhancement with RAG from Internet and local files

• (SOON) Data evaluation & cleaning agent

• (SOON) Open-source model support (+ local inference)

• (SOON) Multimodality

Installation

pip install --upgrade opendatagen

Setting up the OpenAI API key

export OPENAI_API_KEY='your_openai_api_key'

Setting up the SERPLY API key for Google Search API (optional)

export SERPLY_API_KEY='your_serply_api_key'

Usage

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

from opendatagen.data_generator import DataGenerator
from opendatagen.model import LLM
from opendatagen.template import Template, Variable

variation_model = LLM.load_chat.GPT_35_TURBO_CHAT 
completion_model = LLM.load_instruct.GPT_35_TURBO_INSTRUCT

# Create the custom template using the Pydantic models
user_template = Template(
    description="Custom template for Python exercises",
    prompt="Python 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
        )
    }
)

generator = DataGenerator(template=user_template, variation_model=variation_model, completion_model=completion_model)

data = generator.generate_data(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:

from opendatagen.data_generator import DataGenerator
from opendatagen.model import LLM
from opendatagen.template import TemplateManager, TemplateName

variation_model = LLM.load_chat.GPT_35_TURBO_CHAT
completion_model = LLM.load_instruct.GPT_35_TURBO_INSTRUCT

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

generator = DataGenerator(template=template, variation_model=variation_model, completion_model=completion_model)

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

print(data)

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

Contribution

We welcome contributions to Open Datagen! Whether you're looking to fix bugs, add templates, new features, or improve documentation, your help is greatly appreciated.

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.