Self-Refine: Iterative Refinement with Self-Feedback

With Self-Refine, LLMs can generate feedback on their work, use it to improve the output, and repeat this process.

Website | Paper

Table of Contents

Updates
Setup
Getting Started with Acronym Generation
Dialogue Response Generation
Code Readability Improvement
Commongen
GSM-8k
Yelp
PIE
General setup
Citation

Updates

Nov 2023: Added visual self-refine examples and colabs. Use GPT4-V to write tikz code for diagrams, and improve them iteratively.

Stokes' Theorem Example
Visual Self-Refine Example 1

Unicorn Example
Visual Self-Refine Example 2

Setup

We use prompt-lib for querying LLMs. You can install prompt-lib by running the following commands:

git clone https://github.com/reasoning-machines/prompt-lib
pip install prompt-lib/

Depending on your default settings, you may want to use the following to set a PYTHONPATH:

export PYTHONPATH=".:../:.:src:../:../../:.:prompt-lib"

Getting Started with Acronym Generation

python -u src/acronym/run.py "Using language models of code for few-shot commonsense"

Since we sample with temperature 0.7, you'll get different results each time you run the above command.

0 INIT> Using language models of code for few-shot commonsense

0 GEN> CLoCK
0 SCORES> * Ease of pronunciation: CLoCK is pronounced "clah-k". This is easy acronym to pronounce. 4/5 
* Ease of spelling: CLoCK is easy to spell. 5/5 
* Relation to title: CLoCK stands for "Code Language Models of Commonsense Knowledge" which is related to the title. 5/5 
* Positive connotation: CLoCK is a positive acronym. It implies accuracy and efficiency. 5/5 
* Well-known: CLoCK is not a well-known acronym. 2/5

* Total score: 21/25

1 GEN> CMNSC
1 SCORES> * Ease of pronunciation: CMNSC is pronounced "see-min-es-see". This is a difficult acronym to pronounce. 1/5 
* Ease of spelling: CMNSC is difficult to spell. 1/5 
* Relation to title: CMNSC stands for "Commonsense" which is related to the title. 4/5 
* Positive connotation: CMNSC is a neutral acronym. 3/5 
* Well-known: CMNSC is not a well-known acronym. 1/5

* Total score: 10/25

2 GEN> COMMIT
2 SCORES> * Ease of pronunciation: COMMIT is pronounced "com-mit". This is an easy acronym to pronounce. 4/5
* Ease of spelling: COMMIT is easy to spell. 5/5
* Relation to title: COMMIT stands for "Commonsense Model of Code for Intuitive Training" which is related to the title. 5/5 
* Positive connotation: COMMIT is a positive acronym. It implies commitment and dedication. 5/5
* Well-known: COMMIT is not a well-known acronym. 2/5

* Total score: 21/25

Dialogue Response Generation

PYTHONPATH="." python -u src/responsegen/run.py --output <OUTPUT FILE> --size <DATA SIZE>

Use size 0 for running on all test instances

Code Readability Improvement

Note: Please unzip 'data/tasks/codeclean/code_readability/codenet-python-train.jsonl.zip' before running the following commands!

Running:

PYTHONPATH="." python -u src/readability/readability.py --output <OUTPUT FILE>

Evaluation:

PYTHONPATH="." python -u src/readability/{count_comment|count_function|count_meaningful_var}.py --file <INPUT FILE>

Commongen

We use a hard version of commongen. The data is located in data/prompt/commongen. You can download the data by running the following commands:

python -u src/commongen/run.py cmd stair bubble team dryer puppy aliens cat

GSM-8k

To run the GSM-8k task:

python -u src/gsm/run.py

The outputs will be saved in data/tasks/gsm/gsm_outputs.jsonl
To evaluate the outputs:

python src/gsm/gsm_selfref_eval.py --path  data/tasks/gsm/gsm_outputs.jsonl

The evaluation script will also generate a report (data/tasks/gsm/gsm_outputs.jsonl.reports.txt) showing examples of wrong generations, feedback, and refined feedback generations.

Yelp

To run the Yelp task:

python -u src/sentiment_transfer_sr/run.py data/tasks/yelp/yelp-extreme.jso
nl 4 none

The outputs will be saved in data/tasks/yelp/

PIE

To run the PIE task:

python -u src/pie/run.py --slow_programs_file data/tasks/pie/codenet-python-test-1k.jsonl --max_attempts 4 --outfile data/tasks/pie/output --feedback_type rich

For evaluation details, please see docs/pie_eval.md.

General setup

Each task has three different types of prompts:

Init: used to initialize the task. This is how the initial output is generated.
Feedback: used to get feedback from the model on the intermediate results.
Iterate: used to get the next iteration from the model, based on the feedback.

Every task has a run.py that initializes the prompts and runs the task.
As an example, the prompts for commongen are as follows:

Init prompt:

python src/commongen/task_init.py

Feedback prompt:

 python src/commongen/feedback.py

Iterate prompt:

python src/commongen/task_iterate.py

You can also see these prompts on our website.

Citation

@misc{madaan2023selfrefine,
      title={Self-Refine: Iterative Refinement with Self-Feedback}, 
      author={Aman Madaan and Niket Tandon and Prakhar Gupta and Skyler Hallinan and Luyu Gao and Sarah Wiegreffe and Uri Alon and Nouha Dziri and Shrimai Prabhumoye and Yiming Yang and Sean Welleck and Bodhisattwa Prasad Majumder and Shashank Gupta and Amir Yazdanbakhsh and Peter Clark},
      year={2023},
      eprint={2303.17651},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

flowchart LR
    Generator -->|Initializes| Unrefined
    Critic_1 --> Critique_fb
    ... --> Critique_fb
    Critic_k --> Critique_fb
    Critique_fb --> Unrefined{Output to Refine}
    Unrefined --> Refiner
    Refiner --> |R: y_t, x, fb| Refined_Output{Refined Output}
    Refined_Output --> |Stopping Criteria Not Met| Unrefined

madaan / self-refine

readme