Toolkits for discrete, black-box prompt learning based on metaheuristic algorithms.
Simple run the following command to create virtual environment and install all dependencies automatically.
source install.sh
Our experiments are conducted on eight tasks from Natural-Instructions datasets v2 release: task 019, task 021, task 022, task 050, task 069, task 137, task 139, task 195. The datasets can be downloaded from natural-instruction via:
git clone https://github.com/allenai/natural-instructions
Then, you can put it under data path and set your data path with the following parameters:
--data-dir ./data/natural-instructions-2.6/tasks/
Put your Open AI API key in API_KEY_LIST in 'utils/nat_inst_gpt3.py'.
You can run text-babbage-001 + Genetic Algorithm with run.sh:
bash run.shOur code supports both GPT-2 and GPT-3 backbones via the following parameters:
--backbone:gpt2: GPT2-XL (default)gpt3: text-babbage-001 (default)
To use other GPT3 models from the API, please change by the argument --model_name.
--model_name: text-babbage-001 (default), text-ada-001, text-curie-001, text-davinci-001, text-davinci-002, text-davinci-003
We implement several heuristic algorithms to search the optimal discrete prompt for downstream tasks in a black-box way, including Hill Climbing, Simulated Annealing, Genetic Algorithm, Tabu Search and Harmony Search.
These two strategies are based on GrIPS, and we fixed some bugs of their open-sourced code.
For search with Hill Climbing,
python main.py \
--algorithm "hc" \
--mode "Instruction Only" \
--task-idx 0 \
--train-seed 0 \
--num-compose 1 \
--num-candidates 10 \
--num-iter 50 \
--patience 7 \
--write-preds \
--meta-dir "[your_output_dir]" \
--meta-name "HC_bs_20all_edits_l_1_m_10_n_50@task_0_agnostic_trainseed_0_dataseed_42.txt" \
--print-orig \
--agnostic \
--batch-size 20 \
--data-dir "[your_data_path]" \
--project-name 'Plum' \
--checkpoint-freq 10 \
--output-dir "[you_ckpt_dir]"
For search with Simulated Annealing,
python main.py \
--algorithm "hc" \
--mode "Instruction Only" \
--task-idx 0 \
--train-seed 0 \
--num-compose 1 \
--num-candidates 10 \
--num-iter 50 \
--patience 7 \
--write-preds \
--meta-dir "[your_output_dir]" \
--meta-name "SA_bs_20all_edits_l_1_m_10_n_50@task_0_agnostic_trainseed_0_dataseed_42.txt" \
--print-orig \
--agnostic \
--batch-size 20 \
--data-dir "[your_data_path]" \
--project-name 'Plum' \
--checkpoint-freq 10 \
--output-dir "[your_ckpt_dir]" \
--simulated-anneal
We propose a search strategy with the Genetic Algorithm (no crossover) for the optimization of discrete prompts.
For search with Genetic Algorithm,
python main.py \
--algorithm "ga" \
--mode "Instruction Only" \
--task-idx 0 \
--train-seed 0 \
--num-compose 1 \
--num-candidates 10 \
--num-iter 50 \
--patience 7 \
--write-preds \
--meta-dir "[your_output_dir]" \
--meta-name "GA_M_bs_20_all_edits_l_1_m_10_n_50@task_0_agnostic_trainseed_0_dataseed_42_rho_7.txt" \
--print-orig \
--agnostic \
--batch-size 20 \
--tournament-selection 5 \
--data-dir "[your_data_path]" \
--project-name 'Plum' \
--checkpoint-freq 10 \
--output-dir "[your_ckpt_dir]"
The search with Genetic Algorithm also can be combined with Simulated Annealing,
python main.py \
--algorithm "ga" \
--mode "Instruction Only" \
--task-idx 0 \
--train-seed 0 \
--num-compose 1 \
--num-candidates 10 \
--num-iter 50 \
--patience 7 \
--write-preds \
--meta-dir "[your_output_dir]" \
--meta-name "GA_M_bs_20_all_edits_l_1_m_10_n_50@task_0_agnostic_trainseed_0_dataseed_42_rho_7.txt" \
--print-orig \
--agnostic \
--batch-size 20 \
--tournament-selection 5 \
--data-dir "[your_data_path]" \
--project-name 'Plum' \
--checkpoint-freq 10 \
--output-dir "[your_ckpt_dir]" \
--simulated-anneal
We propose a search strategy with the Tabu Search for the optimization of discrete prompts.
For search with Tabu Search,
python main.py \
--algorithm "tabu" \
--mode "Instruction Only" \
--task-idx 0 \
--train-seed 0 \
--num-compose 1 \
--num-candidates 10 \
--num-iter 50 \
--patience 7 \
--write-preds \
--meta-dir "[your_output_dir]" \
--meta-name "Tabu_bs_20_all_edits_l_1_m_10_n_50@task_0_agnostic_trainseed_0_dataseed_42_rho_7.txt" \
--print-orig \
--agnostic \
--batch-size 20 \
--tournament-selection 5 \
--data-dir "[your_data_path]" \
--project-name 'Plum' \
--checkpoint-freq 10 \
--output-dir "[your_ckpt_dir]"
We propose a search strategy with the Harmony Search for the optimization of discrete prompts.
For search with Harmony Search,
python main.py \
--algorithm "hs" \
--mode "Instruction Only" \
--task-idx 0 \
--train-seed 0 \
--num-compose 1 \
--num-candidates 10 \
--num-iter 50 \
--patience 7 \
--write-preds \
--meta-dir "[your_output_dir]" \
--meta-name "HS_bs_20_all_edits_l_1_m_10_n_50@task_0_agnostic_trainseed_0_dataseed_42_rho_7.txt" \
--print-orig \
--agnostic \
--batch-size 20 \
--tournament-selection 5 \
--data-dir "[your_data_path]" \
--project-name 'Plum' \
--checkpoint-freq 10 \
--output-dir "[your_ckpt_dir]"
For help or issues using this package, please submit a GitHub issue.
For personal communication related to this package, please contact Shizhe Diao ([email protected]) and Rui Pan ([email protected]).
We are more than happy if this code is helpful to your work. If you use our code or extend our work, please consider citing our paper:
@article{Plum-bbpl,
title = {Plum: Prompt Learning using Metaheuristic},
author = {Rui Pan, Shuo Xing, Shizhe Diao, Xiang Liu, Kashun Shum, Jipeng Zhang, and Tong Zhang},
year = {2023},
eprint = {2311.08364},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
}