promptzl 0.9.2

Promptzl - LLMs as Classifiers

Pr🥨mptzl v0.9.2 (Beta)

Promptzl is a simple library for turning LLMs into traditional PyTorch-based classifiers using the 🤗 Transformers library.

While large generative models are often used and exhibit strong performance, they can be opaque and slow. Promptzl works in batch mode, returns a softmax distribution, and is 100% transparent. All causal and masked language models from the Hugging Face Hub are available in Promptzl.

Check out more in the official documentation.

Installation

Run

pip install promptzl

or clone this repository, navigate to the main folder and run

pip install .

Getting Started

In just a few lines of code, you can transform a LLM of choice into an old-school classifier with all it's desirable properties:

from promptzl import *
from datasets import load_dataset

dataset = load_dataset("SetFit/ag_news")

verbalizer = Vbz({0: ["World"], 1: ["Sports"], 2: ["Business"], 3: ["Tech"]})
prompt = Txt("[Category:") + verbalizer + Txt("] ") + Key()

model = MaskedLM4Classification("roberta-large", prompt)
output = model.classify(dataset['test'], show_progress_bar=True).predictions
sum([int(prd == lbl) for prd, lbl in zip(output, dataset['test']['label'])]) / len(output)
# 0.7986842105263158

Installation (Dev)

pip install -e .

pip install -r test-requirements.txt

To render the documentation, run:

cd docs && pip install -r requirements.txt && cd ..