RIFF

This repository holds all of the code associated with the research side of my projects considering prompt engineering and paraphrasing the inputs for language models. This repo implements the experiments in our paper:

RIFF: Learning to Rephrase Inputs for Few-shot Fine-tuning of Language Models

https://arxiv.org/pdf/2403.02271

Installing dependencies

Virtualenv installing on linux VM

You can call setup.sh with the OS=linux flag. This installs python in the linux VM and installs the ML libraries for the GPU cards.

bash setup.sh OS=linux ENV_NAME=prompt_torch DEV=true

Virtualenv installing on Vector's GPU Cluster

You can call setup.sh with the OS=vcluster flag. This installs the required libraries to run the experiments on the vector's cluster.

bash setup.sh OS=vcluster ENV_NAME=prompt_torch DEV=true

Using Pre-commit Hooks

The static code checker runs on python3.9.

To check your code at commit time

pre-commit install

You can also get pre-commit to fix your code

pre-commit run

Download data

To download data, use the following link to get 16-shot splits for the classification tasks.

https://github.com/mingkaid/rl-prompt/tree/main/examples/few-shot-classification/data/16-shot

Training Scripts

To train normal models without paraphrase augmentation, use the following train script. Currently, it will be run on a linux machine on a single GPU iteratively for different seeds, prompt optimization techniques and tasks. For a cluster of GPUs managed by slurm, it will submit different jobs in parallel (e.g. for the vcluster)

bash train_scripts/run_basic_experiments.sh

Then you can finetune the paraphraser with the RIFF objective.

bash train_scripts/run_paraphrase_finetuning.sh

Finally, fix the path for the fine-tuned paraphraser, and train the downstream LM with paraphrase augmentation.

bash train_scripts/run_better_experiments.sh

Training Loop

The losses for the downstream language model or the paraphraser model under different optimization techniques have been implemented in the following function:

https://github.com/SaeedNajafi/RIFF/blob/main/src/reference_implementations/prompt_zoo/prompted_lm.py#L754

Reference

If you find the paper or the code helpful, please cite the following article published in ACL 2024 (findings).

@inproceedings{najafi-fyshe-2024-riff,
    title = "{RIFF}: Learning to Rephrase Inputs for Few-shot Fine-tuning of Language Models",
    author = "Najafi, Saeed  and
      Fyshe, Alona",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Findings of the Association for Computational Linguistics ACL 2024",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand and virtual meeting",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-acl.85",
    pages = "1447--1466",
    abstract = "Pre-trained Language Models (PLMs) can be accurately fine-tuned for downstream text processing tasks. Recently, researchers have introduced several parameter-efficient fine-tuning methods that optimize input prompts or adjust a small number of model parameters (e.g LoRA). In this study, we explore the impact of altering the input text of the original task in conjunction with parameter-efficient fine-tuning methods. To most effectively rewrite the input text, we train a few-shot paraphrase model with a Maximum-Marginal Likelihood objective. Using six few-shot text classification datasets, we show that enriching data with paraphrases at train and test time enhances the performance beyond what can be achieved with parameter-efficient fine-tuning alone. The code used for our experiments can be found at https://github.com/SaeedNajafi/RIFF.",
}