GPI: Generative-AI Powered Inference

gpi_pack is a Python library for the statistical inference powered by Generative Artificial Intelligence (AI). It provides a set of tools and utilities for performing statistical inference using the internal representation of the Generative AI models. The library is designed to be easy to use and flexible, allowing users to perform a wide range of statistical analyses.

Note

We released gpi_pack version 0.1.0 on February 27th, 2025. This is the first version of the package, and we currently only support the setting of Text-as-Treatment based on our paper. We have been working hard to make this package as useful and user-friendly as possible. If you have any feedback or suggestions, please feel free to reach out to the maintainer.

Getting Started

• Installation
  • From PyPI
  • From Source
• What’s GPI?
  • Generative-AI Powered Inference
  • Contribute to GPI
• How to use GPU
  • What’s GPU?
  • What if you do not have GPU?
  • Google Colaboratory

Data Generation

• Generating Texts with LLaMa3
  • How to use LLaMa3
  • Creating Texts
  • Repeating Texts
  • Arguments
  • System Prompt
• Generating Texts with Other LLMs
  • Example: Gemma2

Basic Operation

• Text-As-Treatment
  • What is Text-As-Treatment?
  • How to estimate treatment effects
  • How to control confounders
  • Visualizing Propensity Scores
  • Hyperparameters

Advanced Operations

• Hyperparameter Tuning
  • Automated Hyperparameter Tuning
  • List of Hyperparameters
• Customizing Your Analysis
  • TarNet
  • Propensity Score Model
  • Estimation Workflow
• When LLM is too big
  • Model Quantization

References

• dml_score
  • Description
  • Arguments
  • Returns
  • Example Usage
• estimate_k_ate
  • Description
  • Arguments
  • Returns
  • Example Usage
• estimate_psi_split
  • Description
  • Arguments
  • Returns
  • Example Usage
• extract_and_save_hiddens
  • Description
  • Arguments
  • Returns
  • Example Usage
• generate_text
  • Description
  • Arguments
  • Returns
  • Example Usage
• get_instruction
  • Description
  • Arguments
  • Returns
  • Example Usage
• load_hiddens
  • Description
  • Arguments
  • Returns
  • Example Usage
• save_generated_texts
  • Description
  • Arguments
  • Returns
  • Example Usage
• SpectralNormClassifier
  • Description
  • Parameters
  • Example Usage
  • Methods
• TarNet_loss
  • Description
  • Arguments
  • Returns
  • Example Usage
• TarNet
  • Description
  • Parameters
  • Example Usage
  • Methods
• TarNetBase
  • Description
  • Parameters
  • Example Usage
  • Arguments:
• TarNetHyperparameterTuner
  • Description
  • Arguments
  • Example Usage

How to cite

• Imai, Kosuke and Nakamura, Kentaro (2024). Causal Representation Learning with Generative Artificial Intelligence: Application to Texts as Treatments. arXiv preprint arXiv:2410.00903. [Paper]

@article{imai2024causal,
  title={Causal Representation Learning with Generative Artificial Intelligence: Application to Texts as Treatments},
  author={Imai, Kosuke and Nakamura, Kentaro},
  journal={arXiv preprint arXiv:2410.00903},
  year={2024}
}

Copy