Research

This page is a work in progress. More experiments would be added and the page would be prettier in the future.

☂️ BoCoEL

Bayesian Optimization as a Coverage Tool for Evaluating Large Language Models

GitHub License

🤔 Why BoCoEL?

Large language models are expensive and slow behemoths, and evaluating them on gigantic modern datasets only makes it worse.

If only there is a way to just select a meaningful (and small) subset of the corpus and obtain a highly accurate evaluation.....

Wait, sounds like Bayesian Optimization!

Bocoel works in the following steps:

Encode individual entry into embeddings (way cheaper / faster than LLM and reusable).
Use Bayesian optimization to select queries to evaluate.
Use the queries to retrieve from our corpus (with the encoded embeddings).
Profit.

The evaluations generated are easily managed by the provided manager utility.

To our knowledge, this is the first work aiming to reduce computation costs during evaluation (benchmarking) with a (possibly dynamic) budget.

🚀 Features

🎯 Accurately evaluate large language models with just tens of samples from your selected corpus.
💂‍♂️ Uses the power of Bayesian optimization to select an optimal subset of samples for language model to evaluate.
💯 Evalutes the corpus on the model in addition to evaluating the model on corpus.
🤗 Support for GPT2, Pythia, LLAMA and more through integration with huggingface transformers and datasets
🧩 Modular design.
🔎 Efficient representation of the corpus / dataset such as N-sphere representation or whitening of the latent space to agument evaluation quality.

⭐ Give us a star!

Like what you see? Please consider giving this a star (★)!

♾️ Bayesian Optimization

Simply put, Bayesian optimization aims to optimize either the exploration objective (the purple area in the image) or the exploitation object (the height of the black dots). It uses Gaussian processes as a backbone for inference, and uses an acquisition function to decide where to sample next. See here for an a more in-depth introduction.

Since Bayesian optimization works well with expensive-to-evaluate black-box model (paraphrase: LLM), it is perfect for this particular use case. Bocoel uses Bayesian optimization as a backbone for exploring the embedding space given by our corpus, which allows it to select a good subset acting as a mini snapshot of the corpus.

🏎️ Performance Implications

LLMs are painfully slow, especially generative ones (which is what is usually referred to as LLM), since sequence generation is sequential by nature.

Despite bocoel's requirement to use an embedder to encode the entire corpus, embedders are faster than LLMs by orders of magnitude and the time is gained back by practically any savings in evaluating LLMs.

⬇️ Installation

I don't want optional dependencies:

pip install bocoel

Give me the full experience (all optional dependencies):

pip install "bocoel[all]"

🔬 Usage

See the folder examples/getting_started for a simplistic usage of the library to get started with just a few lines of code.

✍️ Develop with BoCoEL

Usage examples are under the folder examples. API reference can be found here.

🥰 Contributing

Contributors wanted! Don't be shy. Feel free to file issues and PRs. For PRs, please follow the guide on contributing and the code of conduct. Openness and inclusiveness are taken very seriously.

🗺️ Roadmap: work in progress

🪑 Simpler usage. I should provide a high level wrapper for the entire library s.t. evaluations can be run in one line.
📊 Visualization module of the evaluation.
🎲 Integration of alternative methods (random, kmedoids...) with Gaussian process.
🥨 Integration with more backends such as VLLM and OpenAI's API.
🆕 Support for Python 3.11+

🏷️ License and Citation

The code is available under Apache License.

If you find this project helpful in your research, please cite this work at

@misc{bocoel2024,
    title = {BoCoEL: Bayesian Optimization as a Coverage Tool for Evaluating Large Language Models},
    url = {https://rentruewang.github.io/bocoel/research/},
    author = {Wang, RenChu},
    month = {January},
    year = {2024}
}

Abstract

BoCoEL, short for Bayesian Optimization as a Coverage Tool for Evaluating Large Language Models, represents an innovative approach in the domain of natural language processing (NLP). This framework leverages Bayesian optimization to efficiently evaluate large language models (LLMs) using a significantly reduced yet representative subset of data from extensive corpora. By encoding the data into embeddings and utilizing Bayesian optimization for sample selection, BoCoEL offers a cost-effective and time-efficient alternative for the evaluation of LLMs. This document delineates the methodology, experimentation, and implications of BoCoEL, highlighting its potential to revolutionize the evaluation process in NLP.

Introduction

The recent proliferation of large language models (LLMs) in NLP has underscored the necessity for efficient evaluation mechanisms. Traditional methods, which involve the assessment of LLMs over vast datasets, are not only time-consuming but also computationally expensive. BoCoEL addresses this challenge by integrating Bayesian optimization into the evaluation process. This approach not only reduces the computational burden but also maintains the integrity and accuracy of the evaluation. The introduction section will further elaborate on the motivation, background, and the specific challenges BoCoEL aims to address in the realm of LLM evaluation.

Methods

The core methodology of BoCoEL revolves around the use of Bayesian Optimization and embeddings to efficiently evaluate large language models (LLMs). This section provides a detailed mathematical overview of these processes.

Embedding Process

The embedding process involves transforming corpus entries into a vector space, facilitating efficient manipulation and comparison. Let \(\mathcal{D}\) be our dataset containing \(N\) entries, \(\mathcal{D} = \{d_1, d_2, ..., d_N\}\). Each entry \(d_i\) is transformed into an embedding vector \(\mathbf{e}_i\) using an embedding function \(E\):

\(\mathbf{e}_i = E(d_i)\)

These embeddings are then used as inputs for the Bayesian Optimization process.

Bayesian Optimization

Bayesian Optimization (BO) is a strategy for global optimization of black-box functions that are expensive to evaluate. It works well with the LLM evaluation problem, as each evaluation can be computationally intensive.

Let \(f: \mathcal{X} \rightarrow \mathbb{R}\) be the expensive black-box function we wish to optimize, where \(\mathcal{X}\) is the space of parameters (in our case, the space of embeddings). BO approximates \(f\) using a surrogate model, typically a Gaussian Process (GP). A GP is defined by its mean function \(m(\mathbf{x})\) and covariance function \(k(\mathbf{x}, \mathbf{x'})\), where \(\mathbf{x}, \mathbf{x'} \in \mathcal{X}\):

\(m(\mathbf{x}) = \mathbb{E}[f(\mathbf{x})]\) \(k(\mathbf{x}, \mathbf{x'}) = \mathbb{E}[(f(\mathbf{x}) - m(\mathbf{x}))(f(\mathbf{x'}) - m(\mathbf{x'}))]\)

The GP posterior distribution after observing data \(\mathcal{D}_n = \{(\mathbf{x}_1, y_1), ..., (\mathbf{x}_n, y_n)\}\) is also a GP:

\(f(\mathbf{x}) | \mathcal{D}_n \sim \mathcal{GP}(\mu_n(\mathbf{x}), \sigma_n^2(\mathbf{x}))\)

where

\(\mu_n(\mathbf{x}) = k_n(\mathbf{x})^T(K_n + \sigma^2I)^{-1}Y_n\) \(\sigma_n^2(\mathbf{x}) = k(\mathbf{x}, \mathbf{x}) - k_n(\mathbf{x})^T(K_n + \sigma^2I)^{-1}k_n(\mathbf{x})\)

with \(k_n(\mathbf{x}) = [k(\mathbf{x}_1, \mathbf{x}), ..., k(\mathbf{x}_n, \mathbf{x})]^T\), \(K_n\) being the covariance matrix formed by applying \(k\) to all pairs of points in \(\mathcal{D}_n\), and \(Y_n = [y_1, ..., y_n]^T\).

An acquisition function \(a: \mathcal{X} \rightarrow \mathbb{R}\) is used to determine where to sample next, balancing exploration and exploitation. Common choices for \(a\) include Expected Improvement (EI) and Upper Confidence Bound (UCB).

\(\text{EI}(\mathbf{x}) = \mathbb{E}[\max(f(\mathbf{x}) - f(\mathbf{x}^+), 0)]\) \(\text{UCB}(\mathbf{x}) = \mu_n(\mathbf{x}) + \kappa \sigma_n(\mathbf{x})\)

where \(\mathbf{x}^+\) is the current best observation, and \(\kappa\) is a parameter controlling the exploration-exploitation trade-off.

Here, an entropy search (ES) scheme is used during our evaluation process. This exploration-focused objective ensures that the bayesian process covers as much of the search space with as few samples as possible, which solves the coverage problem (of the embedding space) effectively.

Using this framework, BoCoEL iteratively selects samples from \(\mathcal{X}\) (the embedding space) to evaluate the LLM, efficiently optimizing the evaluation process.

Experiments

TODO: Add plot TODO: Explain the experiments

The experimental section will present the application of BoCoEL in various scenarios, demonstrating its effectiveness in evaluating different LLMs. Comparative analyses between BoCoEL and traditional evaluation methods will be highlighted, showcasing the efficiency gains in terms of computational resources and time. This section will also include case studies or real-world examples where BoCoEL has been successfully implemented. The results obtained from these experiments will serve to validate the efficacy of the BoCoEL framework in providing accurate evaluations of LLMs using a significantly reduced dataset.

Authors & Acknowledgement

TODO: Add authors.

See the thank you page for details.