ruprompts 0.1.0

Fast prompt tuning framework for large language models

ruPrompts

ruPrompts is a high-level yet extensible library for fast language model tuning via automatic prompt search, featuring integration with HuggingFace Hub, configuration system powered by Hydra, and command line interface.

Prompt is a text instruction for language model, like

Translate English to French:
cat =>

For some tasks the prompt is obvious, but for some it isn't. With ruPrompts you can define only the prompt format, like <P*10>{text}<P*10>, and train it automatically for any task, if you have a training dataset.

You can currently use ruPrompts for text-to-text tasks, such as summarization, detoxification, style transfer, etc., and for styled text generation, as a special case of text-to-text.

Features

• Modular structure for convenient extensibility
• Integration with HF Transformers, support for all models with LM head
• Integration with HF Hub for sharing and loading pretrained prompts
• CLI and configuration system powered by Hydra
• Pretrained prompts for ruGPT-3

Installation

ruPrompts can be installed with pip:

pip install ruprompts[hydra]

See Installation for other installation options.

Usage

Loading a pretrained prompt for styled text generation:

>>> import ruprompts
>>> from transformers import pipeline

>>> ppln_joke = pipeline("text-generation-with-prompt", prompt="konodyuk/prompt_rugpt3large_joke")
>>> ppln_joke("Говорит кружка ложке")
[{"generated_text": 'Говорит кружка ложке: "Не бойся, не утонешь!".'}]

For text2text tasks:

>>> ppln_detox = pipeline("text2text-generation-with-prompt", prompt="konodyuk/prompt_rugpt3large_detox_russe")
>>> ppln_detox("Опять эти тупые дятлы все испортили, чтоб их черти взяли")
[{"generated_text": 'Опять эти люди все испортили'}]

Proceed to Quick Start for a more detailed introduction or start using ruPrompts right now with our Colab Tutorials.

License

ruPrompts is Apache 2.0 licensed. See the LICENSE file for details.