Skip to content

Latest commit

 

History

History
112 lines (79 loc) · 5.8 KB

README.md

File metadata and controls

112 lines (79 loc) · 5.8 KB
Raw

One Prompt is not Enough: Automated Construction of a Mixture-of-Expert Prompts - ICML 2024

Prompt Optimization Mixture-Of-Experts GPT-3.5
Big-Bench-Hard Instruction-Induction Super-Nature-Instruction

This repository contains the official implementation of our MoP, accepted by ICML 2024.

[ICML 2024] One Prompt is not Enough: Automated Construction of a Mixture-of-Expert Prompts
Authors: Ruochen Wang*, Sohyun An*, Minhao Cheng, Tianyi Zhou, Sung Ju Hwang, Cho-Jui Hsieh
Presented by TurningPoint AI - UCLA - UMD - PSU - KAIST

About MoP

Mixture-of-Prompts (MoP) is the first prompt optimization framework that extends the Mixture-of-Expert paradigm to Prompt Optimization. Unlike traditional methods that optimize a single prompt, MoP optimizes a mixture of prompts (experts) with different specialties, each tailored to only handle queries belong to a partition of the entire problem space. In addition, MoP also jointly optimizes the instruction and In-Context demos field of the prompt for each expert. Designed as a plug-and-play solution, MoP is compatible with various prompt generation algorithms, including APE, InstructZero, etc. Our extensive empirical study on 55 tasks across three major benchmarks, one of the largest of its kind, demonstrates that MoP achieves an average win rate of 81% over comparable baselines. Dividing the problem space into sub-regions also benefits the performance on Out-Of-Distribution data.

Overview of MoP pipeline. We adopt the MoE paradigm and divide the problem space into a set of sub-regions. Each sub-region is governed by a specialized expert, equipped with both an instruction and a set of demos. A two-phase process is developed to construct the specialized expert for each region: (1) demo assignment: Inspired by the theoretical connection between ICL and kernel regression, we group demos into experts based on their semantic similarity; (2) instruction assignment: A region-based joint search (RBJS) of an instruction per expert complements the demos assigned to it, yielding a synergistic effect. During inference, each new query is routed to its closest expert in the embedding space and the assigned expert then utilizes its prompt (instruction + demos) to make the final prediction.

Preparations

1. Set environment

All commands are stored in scripts/install_env.sh. If you wish to run InstructZero (iz), run the following installation script as well scripts/install_env_iz.sh.

2. OpenAI account

Our experiments are conducted using gpt-3.5-turbo-instruct model from OpenAI. To use OpenAI API services, you can set the OpenAI Key in scripts/run.sh file.

3. Download benchmarks

Our experiments are conducted on Instruction Induction Benchmark, SuperNature Instruction Benchmark and Big-Bench Hard. Your can download our preprocessed data files HERE, and place them at mop/benchmarks/ (for example mop/benchmarks/superni)

Run MoP

Quick start

Run the following script to verify your setups

bash run-mop-fast.sh
bash run-ape-kcen-fast.sh
bash run-iz-kcen-fast.sh

Instruction Induction Benchmark

bash run.sh --method mop --benchmark ii --datasets [0-24]

Below are the key fields inside run.sh:

  • --method specifies the prompt optimization algorithm.
  • --benchmark is the evaluation benchmark, including ii (Instruction Induction*), sni (SuperNatural Instructions), and ``
  • --datasets is the task_id. You can run on multiple datasets sequentially using comma-separated format (e.g. 1,2,3,4). These benchmarks comprises of hundreds of tasks and it is expensive to evaluate on all of them. We filter around 10-20 tasks from each benchmark, focusing on the challenging ones such as math/coding problems, and those cannot be resolved with a simple prompt (More details in the Appendix).
  • --n_experts is the maximum number of experts. The actual number of expert will be decided by the Kauto algorithm during demo assignment phase.

SuperNature Instruction Benchmark

bash run.sh --method mop --benchmark sni --datasets 0-19

Big-Bench Hard

bash run.sh --method mop --benchmark bbh --datasets 0-9

Run Baselines

To run the baselines, simply change the --method field of the above commands

bash run.sh --method ape
bash run.sh --method ape-random
bash run.sh --method ape-kcen

bash run.sh --method iz
bash run.sh --method iz-random
bash run.sh --method iz-kcen

📖BibTeX

@inproceedings{wang2024one,
  title={One Prompt is not Enough: Automated Construction of a Mixture-of-Expert Prompts},
  author={Wang, Ruochen and An, Sohyun and Cheng, Minhao and Zhou, Tianyi and Hwang, Sung Ju and Hsieh, Cho-Jui},
  booktitle={International Conference on Machine Learning},
  year={2024}
}