Scalable Understanding of Datasets and Models with the Help of Large Language Models

I will make a video tutorial on this topic; stay tuned. This is the library and notebooks to help the audience understand my tutorial.

Installation

I recommend creating a conda environment with python >= 3.9 to use this package.

1. Set up your openai key in your environment. i.e. export OPENAI_API_KEY="[Your OPENAI API KEY]"
2. Installation

Option 1: clone and install locally (for developers)

• git clone git@github.com:ruiqi-zhong/llm_explain.git
• cd llm_explain
• pip install -e .

Option 2: install from github repo

pip3 install --no-cache-dir -v git+https://github.com/ruiqi-zhong/llm_explain.git

Option 3: install from pypi

pip3 install llm-explain

Usage

This repo supports the bare bone implementation for explaining dataset differences and clusters.

See the notebooks, llm_explain/tests/test_cluster.py, and llm_explain/tests/test_diff.py to understand how to use the functions implemented in this repo.

If you want to build on it, refer to other test files to understand the rest of the repo.

A quick example after installation

run python:

>>> from llm_explain.models.diff import explain_diff                                                                                                                           
>>> explain_diff(["cat", "dog", "fish", "carrot", "potato", "apple"], [False, False, False, True, True, True], proposer_num_rounds=2, proposer_num_explanations_per_round=2)

You will get outputs similar to the following in fewer than 30 seconds:

Printing top 3 explanations:
Explanation: refers to a plant-based item; specifically, the text mentions items that grow from plants, including vegetables and fruits. For example, 'carrot' is a type of root vegetable.
Accuracy: 1.0

Explanation: is a type of food; specifically, the text refers to items commonly recognized as food, typically vegetables or fruits. For example, 'apple' is known to be a fruit consumed as food.
Accuracy: 0.8333333333333333

Explanation: mentions a type of food; specifically, the text refers to something that is commonly eaten by humans. For example, 'This apple is very juicy.'
Accuracy: 0.8333333333333333

Notebooks

The notebooks illustrate the following sections in the video tutorial.

• 1.1 Core method: the proposer-validator framework
• 1.1 Extension 1: precise explanations.
• 1.1 Extension 2: goal-constrained explanations.
• 1.1 Extension 3: multiple explanations
• 1.2: Explainable clustering

Related works

related/references.pdf contains the related works mentioned in our presentation. related/main.tex contains the latex source file and references.bib contains the bibtex citations.