jxzhangjhu/sbi_llm

Reliable Gradient-free and Likelihood-free Prompt Tuning

This repository contains official implementation of EACL 2023 (Findings) paper Reliable Gradient-free and Likelihood-free Prompt Tuning.

Prepare your environment

conda create --name SBI_LLM python=3.8
conda activate SBI_LLM
pip install transformers==4.1.1
pip install datasets
pip install fastNLP
pip install cma
pip install sklearn
pip install sbi
pip install uq360

Usage

The main.py file includes all the interfaces required for experiments.

The algorithm.py file implements 4 algorithms proposed in the work:

1. Ensembles: prompt ensembles based on CMA_ES algorithm (for gradient-free prompt tuning).
2. CMA_ELBO: gradient-free variational inference based on CMA_ES algorithm (for gradient-free prompt tuning).
3. ABC_SMC: SBI based algorithm ABC_SMC (for likelihood-free prompt tuning).
4. SBI_neural: neural net-based SBI algorithm (for likelihood-free prompt tuning).

Examples:

To apply ABC_SMC algorithm for likelihood-free prompt tuning on SST2 dataset, and collect 100 prompt samples, simply run:

python main.py \
  --task_name "sst2" \
  --alg_name "ABC_SMC"\
  --num_samples 100 \
  --device "cuda:0" \
  --seed 0 \

To apply Prompt Ensembles algorithm for gradient-free prompt tuning on SNLI dataset, and collect 10 prompt samples, simply run:

python main.py \
  --task_name "snli" \
  --alg_name "Ensembles"\
  --num_samples 10 \
  --device "cuda:0" \
  --seed 0 \

Reference

The implementation is based on this repo: https://github.com/txsun1997/Black-Box-Tuning.

Citations

@inproceedings{shen2023reliable,
  title={Reliable Gradient-free and Likelihood-free Prompt Tuning},
  author={Shen, Maohao and Ghosh, Soumya Sankar and Sattigeri, Prasanna and Das, Subhro and Bu, Yuheng and Wornell, Gregory},
  booktitle={Findings of the Association for Computational Linguistics: EACL 2023},
  pages={2371--2384},
  year={2023}
}