opendatagen 0.0.16

Steerable data generation system for LLM fine-tuning

⬜️ Open Datagen ⬜️

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

Features

• Generate high-quality synthetic datasets using simple templates

• Quality enhancement with RAG from Internet and local files

• Data anonymization

• Data evaluation & cleaning agent

• Open-source model support (HuggingFace Inference API)

• (SOON) Data decontamination

• (SOON) Multimodality

Installation

pip install --upgrade opendatagen

Setting up the OpenAI API key (using openai>=1.2)

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: Generate a low-biased dataset to improve factuality of an LLM.

template.json:

{
    "factuality": {
        "description": "Factuality",
        "prompt": "Given the following text:\n\n'''{wikipedia_content}'''\n\nAnswer to this factually checkable question:\n'''{question}'''.",
        "completion": "Answer: '''{answer}'''. Rate the answer out of 10: {score}",
        "prompt_variation_number": 0,
        "variables": {
            "wikipedia_content": {
                "name": "Wikipedia content",
                "generation_number": 1,
                "get_value_from_huggingface": {
                    "dataset_name": "20220301.en",
                    "dataset_path": "wikipedia",
                    "column_name": "text",
                    "max_tokens": 512
                }
            },
            "question": {
                "name": "Factually checkable question",
                "generation_number": 3,
                "models": [
                    {
                        "openai_chat_model": {
                            "name": "gpt-3.5-turbo-1106",
                            "temperature": 0,
                            "max_tokens": 128
                        }
                    }
                ]
            }, 
            "answer": {
                "name": "Short answer to the question",
                "generation_number": 1,
                "models": [
                    {
                        "openai_instruct_model": {
                            "name": "gpt-3.5-turbo-instruct",
                            "temperature": 0,
                            "max_tokens": 128,
                            "start_with": ["Answer:"]
                        }
                    }
                ]
            
            },
            "score": {
                "name": "Score",
                "generation_number": 1,
                "note": ["You must answer with an integer."],
                "models": [
                    {
                        "openai_chat_model": {
                            "name": "gpt-3.5-turbo-1106",
                            "temperature": 0,
                            "max_tokens": 5
                        }
                    }
                ]
            }
            
        }
    }
}

Python code to generate the dataset:

from opendatagen.template import TemplateManager
from opendatagen.data_generator import DataGenerator

output_path = "factuality.csv"
manager = TemplateManager(template_file_path="template.json")
template = manager.get_template(template_name=template_name)

if template:
    
    generator = DataGenerator(template=template)
    
    data = generator.generate_data(output_path=output_path)
    
    print(data)

Using this template you will:

1. Get text content from the Wikipedia dataset hosted on HuggingFace
2. Generate 3 questions about this content
3. Generate an short answer
4. Rate the answer

Once the CSV created, you can ask an AI Agent to evaluate and correct your dataset

from opendatagen.agent import DataAgent

agent = DataAgent()

agent.run()

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 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.