Here is a demo video explaining the setup and usage:

• UniGen: A Unified Framework for Textual Dataset Generation
  • Demo Video
  • Table of Contents
  • 💎 What's UniGen?
  • 🎯 Features
  • 📕 Usage
    • Installation
    • UniGen Dataset Generation
      • Configuration Settings
        • API Settings
        • Generation Settings
        • Generation Hint
        • Efficiency Configuration
      • Step-by-Step Guide
        • 1. Set Up API Key
        • 2. Prepare the Original Dataset
        • 3. Configure the Generation Settings
        • 4. Define Generation Hints
        • 5. Efficiency Configuration
        • 6. Run Data Generation
    • LLM Evaluation
      • 1. Generate Output
      • 2. Evaluate the Output
    • UniGen Augmentation Using LLaMA-Factory
      • Requirements
      • Fine-Tuning
• Cite UniGen

Large Language Models (LLMs) such as GPT-4 and Llama3 have significantly impacted various fields by enabling high-quality synthetic data generation and reducing dependence on expensive human-generated datasets. Despite this, challenges remain in the areas of generalization, controllability, diversity, and truthfulness within the existing generative frameworks. To address these challenges, this paper presents UniGen, a comprehensive LLM-powered framework designed to produce diverse, accurate, and highly controllable datasets. UniGen is adaptable, supporting all types of text datasets and enhancing the generative process through innovative mechanisms. To augment data diversity, UniGen incorporates an attribute-guided generation module and a group checking feature. For accuracy, it employs a code-based mathematical assessment for label verification alongside a retrieval-augmented generation technique for factual validation. The framework also allows for user-specified constraints, enabling customization of the data generation process to suit particular requirements.

TL;DR: UniGen is an LLM-powered framework designed to generate diverse, accurate, and highly controllable text datasets.

• Generalization: UniGen supports all textual datasets as input to generate a new dataset.
• Diversity: We support Attribute-Guided Generation, Diverse Example Selection for ICL, and Group Checking to enhance data diversity.
• Truthfulness: Self-Evaluation, Code-Based Validation, and RAG-Based Validation are equipped to ensure truthfulness.
• Controllability: UniGen accepts user constraints to make generation more controllable.
• Various Application: UniGen can be applied for dynamic benchmark or data augmentation.

git clone --depth 1 [email protected]:HowieHwong/UniGen.git
cd UniGen
pip install -e .

This guide provides detailed instructions for generating a dataset using the UniGen tool with the specified configuration settings.

Below are the configurations to be used for data generation:

api_settings:
  model_type: "gpt"
  api_key: "YOUR_API_KEY_HERE"
  embedding_model: "text-embedding-3-large"
  use_azure: false
  api_base: null
  azure_version: null

• model_type: Specifies the type of model to use (e.g., GPT).
• api_key: Your API key for accessing the model.
• embedding_model: The embedding model to use.
• use_azure: Boolean to indicate if Azure should be used.
• api_base: Base URL for the API (if using Azure).
• azure_version: Version of Azure to use (if applicable).

generation_settings:
  temperature: 0.8
  generation_number: 100
  batch_size: 5
  few_shot_num: 5
  random_example: false
  max_worker: 2

• temperature: Controls the randomness of the generation.
• generation_number: Total number of items to generate.
• batch_size: Number of items to generate in each batch.
• few_shot_num: Number of examples to use for few-shot learning.
• random_example: Boolean to indicate if random examples should be used.
• max_worker: Number of worker threads to use for generation.

generation_hint:
  dataset_description: "It is a dataset of high quality linguistically diverse grade school math word problems created by human problem writers. These problems take between 2 and 8 steps to solve, and solutions primarily involve performing a sequence of elementary calculations using basic arithmetic operations (+ − ×÷) to reach the final answer. A bright middle school student should be able to solve every problem. It can be used for multi-step mathematical reasoning. Each problem should only have one question and one correct answer."
  dataset_name: "GSM8K"
  original_dataset_path: "path/to/original/dataset"
  dataset_constraint: []
  with_label: true
  with_attribute: false
  add_attribute: false
  extra_info_keys: []

• dataset_description: A detailed description of the dataset.
• dataset_name: The name of the dataset.
• original_dataset_path: Path to the original dataset.
• dataset_constraint: List of constraints for generating the dataset.
• with_label: Boolean to indicate if labels should be included.
• with_attribute: Boolean to indicate if attributes should be included.
• add_attribute: Boolean to indicate if new attributes should be added.
• extra_info_keys: List of extra information keys.

efficiency_configuration:
  learn_from_mistake: false
  learn_from_human_feedback: false
  feedback_iteration: 1
  self_reflection: true
  math_eval: true
  truthfulness_eval: false

• learn_from_mistake: Boolean to indicate if the model should learn from mistakes.
• learn_from_human_feedback: Boolean to indicate if the model should learn from human feedback.
• feedback_iteration: Number of feedback iterations.
• self_reflection: Boolean to indicate if self-reflection should be used.
• math_eval: Boolean to indicate if mathematical evaluation should be used.
• truthfulness_eval: Boolean to indicate if truthfulness evaluation should be used.

Replace "YOUR_API_KEY_HERE" in the api_settings with your actual API key.

Ensure that the original dataset is available at the specified path in original_dataset_path.

Adjust the generation_settings according to your specific needs. Ensure the values align with your data generation goals.

Provide detailed information in generation_hint to guide the data generation process. This helps in creating high-quality and relevant data.

Configure efficiency settings to optimize the data generation process. Enable or disable features like learning from mistakes and self-reflection based on your requirements.

Use the configured settings to generate the dataset using the UniGen tool. Ensure all configurations are correctly set before starting the generation process.

unigen-cli gene examples/eval_generation.yaml

To assess the performance of LLMs on the generated dataset, follow these steps:

First, generate the output using the provided dataset. Execute the following command to generate the output for the LLMs on your dataset:

unigen-cli evaluation examples/eval_config.yaml

You can customize your evaluation settings in the configuration file located at examples/eval_generation_config.yaml.

After generating the output, proceed to evaluate its performance. Use the following command to evaluate the generated output:

unigen-cli judge examples/eval_config.yaml

Customize your evaluation settings in the configuration file located at examples/eval_judge.yaml.

Specify the file to be evaluated in the tasks_files section of your configuration. Include the filename and the temperature setting for the LLM-judge process.

This section aims to introduce how to use the generated data powered by UniGen to conduct data augmentation on your LLMs.

• LLaMA-Factory installed

git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e .

To use LLaMA-Factory, follow these steps:

1. Add your dataset into dataset_info.json.

   • The dataset_info.json file should include details about your dataset. For more information, refer to the LLaMA-Factory data documentation.

2. Update the train_config.yaml file with your dataset information and training parameters.

3. Start the training process with the following command:

   llamafactory-cli train train_config.yaml

4. Once training is complete, you can run model inference using the command:

   llamafactory-cli api model_inference.yaml

@misc{wu2024unigenunifiedframeworktextual,
      title={UniGen: A Unified Framework for Textual Dataset Generation Using Large Language Models}, 
      author={Siyuan Wu and Yue Huang and Chujie Gao and Dongping Chen and Qihui Zhang and Yao Wan and Tianyi Zhou and Xiangliang Zhang and Jianfeng Gao and Chaowei Xiao and Lichao Sun},
      year={2024},
      eprint={2406.18966},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.18966}, 
}