Skip to content

An example of how to fine-tune a large language model with prefix-tuning (Li & Liang 2021), with machine translation as the use case.

Notifications You must be signed in to change notification settings

SullyChen/Prefix-Tuning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 

Repository files navigation

Prefix Fine-tuning

This notebook demonstrates an example of prefix tuning, where learned embeddings allow for a comparatively low-memory method to fine-tune behavior in a large language model via encoding behavior/information in a small number of embedding tokens. This notebook demonstrates an example of fine-tuning GPT-2 XL to translate from French to English and vice versa. The way this works is essentially as follows: we learn two prefixes, [start_english] and [start_french]. When we want to translate a French sentence to English, we would input the following prompt to the language model:

[start_french]Bonjour! Je m'appelle Sully.[start_english].

We can also do the same task via few-shot prompting of the language model. At the end of this notebook, we compare (via BLEU-score) the performance of prefix-tuning with few-shot translation and show a marked improvement. Ideally, we would do a comparison of fine-tuning the whole model on the dataset vs. prefix-tuning, but I don't have enough GPU memory on my machine to do that test. Anyway, we see a ~2x improvement in BLEU score via prefix-tuning versus 3-shot prompting!

Requirements

  • Dataset (we won't use all of the dataset, only ~200k examples of the dataset for training and evaluation. If you don't have enough memory to load the whole dataset, just load a subset instead!)
  • Pytorch (for machine learning)
  • NTLK (for BLEU score)
  • Numpy (for math)
  • TQDM (for progress bars)
  • HuggingFace transformers (for GPT-2 XL)

Please see the Jupyter notebook in the repository for a detailed explanation and write-up!

Example

Original French:

Cela est fait, en partie, pour répondre aux besoins des sénateurs, qui sont souvent membres de plusieurs comités.

Ground Truth English:

This is done, in part, to better accommodate the needs of senators who are often members of several committees.

GPT-2 XL 3-shot:

The program is designed to help participants become more involved in their communities.

GPT-2 XL prefix-tuning:

This is done for the benefit of the participants, who are members of many different organizations.

About

An example of how to fine-tune a large language model with prefix-tuning (Li & Liang 2021), with machine translation as the use case.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published